Seungmoon ChoichoispurdueeduHaptic Interface ResearchLaboratoryPurdue UniversityElectrical Engineering Building465 Northwestern AvenueWest Lafayette IN 47907

Hong Z TanhongtanpurdueeduHaptic Interface ResearchLaboratoryPurdue UniversityElectrical Engineering Building465 Northwestern AvenueWest Lafayette IN 47907

Presence Vol 13 No 4 August 2004 395ndash415

copy 2004 by the Massachusetts Institute of Technology

Perceived Instability of VirtualHaptic Texture I ExperimentalStudies

Abstract

This paper presents a quantitative characterization of the instability that a humanuser often experiences while interacting with a virtual textured surface renderedwith a force-reflecting haptic interface First we quantified the degree of stabilityinstability during haptic texture rendering through psychophysical experiments Thestiffness of the virtual textured surface upon detection of instability was measuredunder a variety of experimental conditions using two texture rendering methodstwo exploration modes and various texture model parameters We found that therange of stiffness values for stable texture rendering was quite limited Second weinvestigated the attributes of the proximal stimuli experienced by a human handwhile exploring the virtual textured surface in an attempt to identify the sources ofperceived instability Position force and acceleration were measured and then ana-lyzed in the frequency domain The results were characterized by sensation levels interms of spectral intensity in dB relative to the human detection threshold at thesame frequency We found that the spectral bands responsible for texture and in-stability perception were well separated in frequency such that they excited differ-ent mechanoreceptors and were therefore perceptually distinctive Furthermorewe identified the high-frequency dynamics of the device to be a likely source ofperceived instability Our work has implications for displaying textured surfacesthrough a force feedback device in a virtual environment

1 Introduction

Haptic texture rendering is a growing research field that holds muchpromise for enriching the sensory attributes of objects in a virtual environmentand for allowing precise and systematic control of textured surfaces for psycho-physical studies Our long-term research objectives are to have a better under-standing of how to characterize surface textures in physical and perceptualspaces and to develop procedure-based rendering algorithms that can effec-tively span the human perceptual space for haptic texture The first problemthat we have encountered is the instability perceived by a human user whileinteracting with a virtual textured surface rendered with a force-reflecting hap-tic interface This problem has also been reported anecdotally by other re-searchers (see for example Wall amp Harwin 2000 Weisenberger Krier ampRinker 2000) Given the increasingly popular use of force feedback devices forrendering virtual textures it is imperative that we have a better understandingof the conditions under which haptic virtual textures are free of any unin-

Choi and Tan 395

tended perceptual artifacts We use the term ldquoperceivedinstabilityrdquo to denote any unrealistic sensations (such asbuzzing or apparent aliveness of a surface) that can notbe attributed to the physical properties of the textureunder examination

Although everyone has some notion of what textureis the concept of texture is not clearly defined Katzconsidered texture as the fine structure of a surface (mi-crogeometry) and as independent of the shape (macro-geometry) of an object or surface (Katz 19251989)The systematic study of haptic texture perception beganabout thirty years ago (Lederman amp Taylor 1972) Onetopic that has been controversial is whether informationabout surface texture is encoded spatially or temporallyBoth types of information are available during direct(fingerpad) exploration but only temporal cues (vibra-tions) are available during indirect (probe-mediated)exploration The consensus that has emerged from psy-chophysical and neurophysiological studies is that hu-mans use temporal cues while exploring surface texturesvia a probe While the same temporal cues are availableduring fingerpad exploration humans prefer to use in-tensive (depth of microstructures) andor spatial (sizeof microstructures) cues instead (Johnson amp Hsiao1992) Performance with a bare fingerpad was better fortasks requiring spatial judgments such as haptic objectrecognition but roughness perception was very similarwhether the direct or the indirect method was used (Le-derman amp Klatzky 1999) Recently Lederman and hercolleagues have found that exploration speed has a sub-stantial effect on texture perception thereby supportinga theory based on temporal coding of texture (Leder-man Klatzky Hamilton amp Ramsay 1999) In addi-tion neurophysiological and psychophysical data sug-gest that temporal cues are responsible for perception ofvery fine surface details (with inter-element spacing be-low 1 mm) (LaMotte amp Srinivasan 1991 Johnson ampHsiao 1994 Hollins amp Risner 2000) For very smoothsurfaces the probe method produced greater perceivedroughness than the fingerpad method (Klatzky amp Le-derman 1999) Therefore probe-mediated surface tex-ture perception should yield results similar to the directmethod with better performance expected for verysmall-scale (less than 1 mm) surface features

From the above discussions it is evident that force-feedback devices that emulate probe-mediated textureexploration should produce successful rendering of tex-tured surfaces Indeed the development of computa-tional methods for texture rendering that is virtual en-vironment dynamics models for textures has receivedincreased attention from the haptics research commu-nity in the past few years Minskyrsquos Sandpaper systemwas perhaps the first successful attempt at generatingsynthetic textures (Minsky 1995 Minsky amp Lederman1996) Using a two degree-of-freedom (DoF) force-reflecting joystick Minsky developed a tangential force-gradient algorithm for 2D texture rendering where thedisplayed force was in the plane of the textured surfaceand proportional to the gradient of the surface-heightprofile Several successful implementations of texturerendering methods using three (or more) DoF force-reflecting devices have also been reported For 3D hap-tic rendering resistive forces are rendered to prevent thepenetration of an interaction tool into the objectsthereby conveying the shape of virtual objects To add asense of surface texture variations are imposed on thesebaseline resistive forces Massie reported that changingthe magnitude of the resistive forces alone can generatethe perception of textures (Massie 1996) Ho Bas-dogan and Srinivasan (1999) developed more sophisti-cated texture rendering algorithms by using the bumpmapping technique in computer graphics to add pertur-bations to both the magnitude and direction of the re-sistive forces for various texture models Other research-ers have applied stochastic texture models (Fritz ampBarner 1996 Siira amp Pai 1996 Costa amp Cutkosky2000 Kim Kyrikou Sukhatme amp Desbrun 2002) andvibration-based models to haptic texture rendering(Okamura Dennerlein amp Howe 1998) Most studieson haptic texture rendering have focused on the devel-opment of efficient algorithms that are fast enough(with an update rate of 1 kHz or higher) for haptic ren-dering These algorithms can potentially produce virtualtextures with well-controlled surface characteristics andthus lend themselves well to perceptual studies on tex-ture To the best of our knowledge however few stud-ies have investigated the conditions under which hapti-

396 PRESENCE VOLUME 13 NUMBER 4

cally rendered textures are guaranteed to be perceptuallyldquocleanrdquo (ie free of artifacts)

In general there are two major sources of perceivedinstability during haptic texture rendering an improperenvironment model and unstable control of the hapticinterface Realistic texture rendering requires the envi-ronment model to faithfully follow the physical proper-ties of the objects being rendered In practice howeverthe environment model is usually an approximation tothe underlying physics with the goal of inducing a tar-get percept such as the roughness of a textured surfaceTo the extent that such simplification induces percep-tual artifacts the user can perceive a virtual surface to beunrealistic or unstable The stability of the haptic inter-face in a control sense is also a necessary condition forrealistic rendering Perception of buzzing and chatteringare likely caused by control instability Unbounded be-havior of a controller unmodeled dynamics of a hapticinterface quantization noise of encoders energy instill-ing effects of a zero-order-hold converter and asyn-chronous switch time can all lead to control instability(Gillespie amp Cutkosky 1996) Although many controltheoretical studies have tackled the stability problem ofvirtual walls using a simplified one-DoF interactionmodel (Gillespie amp Cutkosky 1996 Adams amp Han-naford 1999 Miller Colgate amp Freeman 2000 Han-naford Ryu amp Kim 2001 Hannaford amp Ryu 2002)these techniques cannot be easily extended to the inves-tigation of the stability of haptic texture rendering Do-ing so requires modeling haptic interaction as a multi-DoF problem and extending the surface model from asingle flat surface to complex curved surfaces therebysignificantly increasing the complexity of the analysis

Any study of perceived instability has to take into ac-count the effect of both environment modeling andcontrol stability In addition the human decision pro-cess also plays a role in the perceived quality of virtualtextures The experiments reported in this paper weredesigned to address the following two questions withhuman subjects

Under what conditions do human users perceiveinstability from virtual haptic textures rendered witha force feedback device

What are the proximal stimuli that are responsiblefor the perception of instability and why

In the first set of experiments we quantified the max-imum rendering stiffness under which virtual texturesare perceived to be realistic We conducted psychophysi-cal experiments using various texture model parameterstexture rendering methods and exploration modesQualitative descriptions of different kinds of instabilitywere also collected in these experiments In the secondset of experiments we identified the proximal stimulithat caused the perception of instability This was ac-complished through measurements and analyses of thesignals recorded on the stylus that interfaced the humanhand with the haptic interface Some of the results havebeen published earlier in Choi and Tan (2002a 2002b)

The remainder of this paper is organized as followsThe methods that are common to both psychophysicaland measurement experiments are laid out in section 2The specific methods and results for the psychophysicalexperiments are presented in section 3 Those for mea-surement experiments can be found in section 4 Weconclude the paper with a general discussion in section 5

2 Rendering and Exploration of VirtualTextures Used in Our Experiments

In this section we provide the details of the tex-ture model rendering methods and exploration modesused in both psychophysical and measurement experi-ments

21 Texture Model

The virtual textured surfaces were modeled as 1Dsinusoidal gratings superimposed on a flat surface Thisunderlying flat surface defined by z 0 in the worldcoordinate frame of the PHANToM formed a verticalwall facing the user of the PHANToM (see Figure 1)The sinusoidal grating was described by

z A sin ([2L]x) A

where A and L denote the amplitude and (spatial) wave-

Choi and Tan 397

length respectively (see Figure 2) Sinusoidal gratingshave been widely used as the basic building blocks fortextured surfaces for studies on haptic texture percep-tion (Lederman et al 1999 Weisenberger et al 2000)They have also been used as the basis of a function setfor modeling real haptic textures (Wall amp Harwin1999)

22 Texture Rendering Method

Two basic texture rendering methods were em-ployed in the current study Both methods use a spring

model to calculate the magnitude of the rendered forceas K d(t) where K is the stiffness of the textured sur-face and d(t) is the penetration depth of the stylus attime t (see Figure 2) The penetration depth is calcu-lated as follows

dt 0 if pzt 0A sin 2 pxtL A pzt if pzt 0 (1)

where p(t) (px(t) py(t) pz(t)) is the position of the tipof the stylus

The two methods differ in the way the force direc-tions are rendered The first method introduced byMassie (1996) renders a force Fmag(t) with a constantdirection normal to the underlying flat wall of the tex-tured surface The second method proposed by Ho etal (1999) renders a force Fvec(t) with varying directionssuch that it remains normal to the local microgeometryof the sinusoidal texture model Mathematically

Fmagt Kdt nW (2)

Fvect Kdt nT pt (3)

where nW is the normal vector of the underlying flatwall and nT(p(t)) is the normal vector of the texturedsurface at p(t) Both methods keep the force vectors inthe horizontal plane (zx plane in Figure 1) therebyeliminating the effect of gravity on rendered forces

The two texture rendering methods are natural exten-sions of virtual flat wall rendering techniques Perceptu-ally they are very different textures rendered by Fvec(t)feel rougher than those rendered by Fmag(t) with thesame texture model Textures rendered by Fvec(t) canalso feel sticky sometimes

23 Exploration Mode

An exploration mode refers to a stereotypical pat-tern of the motions that humans employ to perceive acertain object attribute through haptic interaction (Le-derman amp Klatzky 1987) In our experiments we testedtwo exploration modes free exploration and strokingto examine the effect of user interaction patterns on in-stability perception In the free exploration mode sub-jects were allowed to use the interaction pattern that

Figure 1 An illustration of the virtual textured surfaces and the two

coordinate frames used in our experiments Position of the stylus tip

was always measured in the world coordinate frame

Figure 2 An illustration of the parameters used in texture

rendering

398 PRESENCE VOLUME 13 NUMBER 4

they found most effective at discovering instability ofvirtual textures This mode was selected to be the mostchallenging interaction pattern for a haptic texture ren-dering system in terms of perceived instability In thestroking mode subjects were instructed to move thestylus laterally across the textured surface (ie along thex axis as shown in Figure 1) Stroking is the explorationmode most frequently employed by humans for textureperception and identification (Lederman amp Klatzky1987)

3 Psychophysical Experiments

In this section we describe the experimental de-sign and results of the psychophysical experiments con-ducted for a quantitative analysis of perceived instabilityduring haptic texture rendering The specific objectivesof these experiments were (1) to quantify the parameterspace that results in virtual textures that are perceived tobe stable and (2) to gather qualitative descriptions ofdifferent kinds of perceived instability

31 Experiment Design

311 Apparatus A PHANToM force-reflectinghaptic interface (model 10A SensAble TechnologiesWoburn MA) was used in all experiments to rendervirtual textured surfaces It was equipped with a stylus asan interaction tool and encoder gimbals for orientationsensing This PHANToM model has a nominal posi-tional resolution of 003 mm and a nominal maximumstiffness of 35 Nmm

312 Stimuli The virtual texture models used inthe experiments had two parameters amplitude A andwavelength L of the 1D sinusoidal gratings Three val-ues of A (05 10 and 20 mm) and three values of L(10 20 and 40 mm) were tested resulting in a totalof nine textured surface profiles These sinusoidal grat-ings can be well constructed due to the 003 mm nomi-nal positional resolution of the PHANToM Each sur-face profile was rendered with the two texture renderingmethods described in Equations 2 and 3 with stiffness

K as the parameter It follows that the three parametersA L and K along with the texture rendering methodsuniquely defined the stimuli used in this study

Due to the fact that the PHANToM workspaceboundary exhibits inferior dynamics performance thevirtual textured surface was restricted to a 15 cm 15cm region located near the center of the PHANToMworkspace

313 Subjects Three subjects participated inthese experiments One subject (S1 male) was an expe-rienced user of the PHANToM haptic interface Theother two subjects (S2 and S3 females) had not usedany haptic interface prior to this study The average ageof the subjects was 263 years old All subjects are right-handed and reported no known sensory or motor ab-normalities with their upper extremities

314 Conditions The independent variablesemployed in the experiments were texture renderingmethod exploration mode and amplitude and wave-length values of sinusoidal surface profiles Four experi-ments defined by the combinations of the two texturerendering methods and the two exploration modeswere conducted There were nine conditions (3 A 3L values) per experiment (see Table 1)

The dependent variable measured in all 36 experi-mental conditions (4 experiments 9 conditions perexperiment) was the maximum stiffness KT below whichthe rendered textured surface was perceived to be stable

315 Procedure All subjects went through ini-tial training to develop criteria for the perception of in-stability of a virtual textured surface During the train-ing the subject chose the texture rendering method andselected the values of A L and K The subject was in-formed that the virtual textures were rendered as 1Dsinusoidal gratings The subject was instructed to regardany sensation that felt unrealistic based on his or herexperience of real textures as an indication of perceivedinstability Each subject spent approximately one houron training

The method of limits (Gescheider 1985) was used inall experiments Given a pair of A and L values within

Choi and Tan 399

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

tended perceptual artifacts We use the term ldquoperceivedinstabilityrdquo to denote any unrealistic sensations (such asbuzzing or apparent aliveness of a surface) that can notbe attributed to the physical properties of the textureunder examination

Although everyone has some notion of what textureis the concept of texture is not clearly defined Katzconsidered texture as the fine structure of a surface (mi-crogeometry) and as independent of the shape (macro-geometry) of an object or surface (Katz 19251989)The systematic study of haptic texture perception beganabout thirty years ago (Lederman amp Taylor 1972) Onetopic that has been controversial is whether informationabout surface texture is encoded spatially or temporallyBoth types of information are available during direct(fingerpad) exploration but only temporal cues (vibra-tions) are available during indirect (probe-mediated)exploration The consensus that has emerged from psy-chophysical and neurophysiological studies is that hu-mans use temporal cues while exploring surface texturesvia a probe While the same temporal cues are availableduring fingerpad exploration humans prefer to use in-tensive (depth of microstructures) andor spatial (sizeof microstructures) cues instead (Johnson amp Hsiao1992) Performance with a bare fingerpad was better fortasks requiring spatial judgments such as haptic objectrecognition but roughness perception was very similarwhether the direct or the indirect method was used (Le-derman amp Klatzky 1999) Recently Lederman and hercolleagues have found that exploration speed has a sub-stantial effect on texture perception thereby supportinga theory based on temporal coding of texture (Leder-man Klatzky Hamilton amp Ramsay 1999) In addi-tion neurophysiological and psychophysical data sug-gest that temporal cues are responsible for perception ofvery fine surface details (with inter-element spacing be-low 1 mm) (LaMotte amp Srinivasan 1991 Johnson ampHsiao 1994 Hollins amp Risner 2000) For very smoothsurfaces the probe method produced greater perceivedroughness than the fingerpad method (Klatzky amp Le-derman 1999) Therefore probe-mediated surface tex-ture perception should yield results similar to the directmethod with better performance expected for verysmall-scale (less than 1 mm) surface features

From the above discussions it is evident that force-feedback devices that emulate probe-mediated textureexploration should produce successful rendering of tex-tured surfaces Indeed the development of computa-tional methods for texture rendering that is virtual en-vironment dynamics models for textures has receivedincreased attention from the haptics research commu-nity in the past few years Minskyrsquos Sandpaper systemwas perhaps the first successful attempt at generatingsynthetic textures (Minsky 1995 Minsky amp Lederman1996) Using a two degree-of-freedom (DoF) force-reflecting joystick Minsky developed a tangential force-gradient algorithm for 2D texture rendering where thedisplayed force was in the plane of the textured surfaceand proportional to the gradient of the surface-heightprofile Several successful implementations of texturerendering methods using three (or more) DoF force-reflecting devices have also been reported For 3D hap-tic rendering resistive forces are rendered to prevent thepenetration of an interaction tool into the objectsthereby conveying the shape of virtual objects To add asense of surface texture variations are imposed on thesebaseline resistive forces Massie reported that changingthe magnitude of the resistive forces alone can generatethe perception of textures (Massie 1996) Ho Bas-dogan and Srinivasan (1999) developed more sophisti-cated texture rendering algorithms by using the bumpmapping technique in computer graphics to add pertur-bations to both the magnitude and direction of the re-sistive forces for various texture models Other research-ers have applied stochastic texture models (Fritz ampBarner 1996 Siira amp Pai 1996 Costa amp Cutkosky2000 Kim Kyrikou Sukhatme amp Desbrun 2002) andvibration-based models to haptic texture rendering(Okamura Dennerlein amp Howe 1998) Most studieson haptic texture rendering have focused on the devel-opment of efficient algorithms that are fast enough(with an update rate of 1 kHz or higher) for haptic ren-dering These algorithms can potentially produce virtualtextures with well-controlled surface characteristics andthus lend themselves well to perceptual studies on tex-ture To the best of our knowledge however few stud-ies have investigated the conditions under which hapti-

396 PRESENCE VOLUME 13 NUMBER 4

cally rendered textures are guaranteed to be perceptuallyldquocleanrdquo (ie free of artifacts)

In general there are two major sources of perceivedinstability during haptic texture rendering an improperenvironment model and unstable control of the hapticinterface Realistic texture rendering requires the envi-ronment model to faithfully follow the physical proper-ties of the objects being rendered In practice howeverthe environment model is usually an approximation tothe underlying physics with the goal of inducing a tar-get percept such as the roughness of a textured surfaceTo the extent that such simplification induces percep-tual artifacts the user can perceive a virtual surface to beunrealistic or unstable The stability of the haptic inter-face in a control sense is also a necessary condition forrealistic rendering Perception of buzzing and chatteringare likely caused by control instability Unbounded be-havior of a controller unmodeled dynamics of a hapticinterface quantization noise of encoders energy instill-ing effects of a zero-order-hold converter and asyn-chronous switch time can all lead to control instability(Gillespie amp Cutkosky 1996) Although many controltheoretical studies have tackled the stability problem ofvirtual walls using a simplified one-DoF interactionmodel (Gillespie amp Cutkosky 1996 Adams amp Han-naford 1999 Miller Colgate amp Freeman 2000 Han-naford Ryu amp Kim 2001 Hannaford amp Ryu 2002)these techniques cannot be easily extended to the inves-tigation of the stability of haptic texture rendering Do-ing so requires modeling haptic interaction as a multi-DoF problem and extending the surface model from asingle flat surface to complex curved surfaces therebysignificantly increasing the complexity of the analysis

Any study of perceived instability has to take into ac-count the effect of both environment modeling andcontrol stability In addition the human decision pro-cess also plays a role in the perceived quality of virtualtextures The experiments reported in this paper weredesigned to address the following two questions withhuman subjects

Under what conditions do human users perceiveinstability from virtual haptic textures rendered witha force feedback device

What are the proximal stimuli that are responsiblefor the perception of instability and why

In the first set of experiments we quantified the max-imum rendering stiffness under which virtual texturesare perceived to be realistic We conducted psychophysi-cal experiments using various texture model parameterstexture rendering methods and exploration modesQualitative descriptions of different kinds of instabilitywere also collected in these experiments In the secondset of experiments we identified the proximal stimulithat caused the perception of instability This was ac-complished through measurements and analyses of thesignals recorded on the stylus that interfaced the humanhand with the haptic interface Some of the results havebeen published earlier in Choi and Tan (2002a 2002b)

The remainder of this paper is organized as followsThe methods that are common to both psychophysicaland measurement experiments are laid out in section 2The specific methods and results for the psychophysicalexperiments are presented in section 3 Those for mea-surement experiments can be found in section 4 Weconclude the paper with a general discussion in section 5

2 Rendering and Exploration of VirtualTextures Used in Our Experiments

In this section we provide the details of the tex-ture model rendering methods and exploration modesused in both psychophysical and measurement experi-ments

21 Texture Model

The virtual textured surfaces were modeled as 1Dsinusoidal gratings superimposed on a flat surface Thisunderlying flat surface defined by z 0 in the worldcoordinate frame of the PHANToM formed a verticalwall facing the user of the PHANToM (see Figure 1)The sinusoidal grating was described by

z A sin ([2L]x) A

where A and L denote the amplitude and (spatial) wave-

Choi and Tan 397

length respectively (see Figure 2) Sinusoidal gratingshave been widely used as the basic building blocks fortextured surfaces for studies on haptic texture percep-tion (Lederman et al 1999 Weisenberger et al 2000)They have also been used as the basis of a function setfor modeling real haptic textures (Wall amp Harwin1999)

22 Texture Rendering Method

Two basic texture rendering methods were em-ployed in the current study Both methods use a spring

model to calculate the magnitude of the rendered forceas K d(t) where K is the stiffness of the textured sur-face and d(t) is the penetration depth of the stylus attime t (see Figure 2) The penetration depth is calcu-lated as follows

dt 0 if pzt 0A sin 2 pxtL A pzt if pzt 0 (1)

where p(t) (px(t) py(t) pz(t)) is the position of the tipof the stylus

The two methods differ in the way the force direc-tions are rendered The first method introduced byMassie (1996) renders a force Fmag(t) with a constantdirection normal to the underlying flat wall of the tex-tured surface The second method proposed by Ho etal (1999) renders a force Fvec(t) with varying directionssuch that it remains normal to the local microgeometryof the sinusoidal texture model Mathematically

Fmagt Kdt nW (2)

Fvect Kdt nT pt (3)

where nW is the normal vector of the underlying flatwall and nT(p(t)) is the normal vector of the texturedsurface at p(t) Both methods keep the force vectors inthe horizontal plane (zx plane in Figure 1) therebyeliminating the effect of gravity on rendered forces

The two texture rendering methods are natural exten-sions of virtual flat wall rendering techniques Perceptu-ally they are very different textures rendered by Fvec(t)feel rougher than those rendered by Fmag(t) with thesame texture model Textures rendered by Fvec(t) canalso feel sticky sometimes

23 Exploration Mode

An exploration mode refers to a stereotypical pat-tern of the motions that humans employ to perceive acertain object attribute through haptic interaction (Le-derman amp Klatzky 1987) In our experiments we testedtwo exploration modes free exploration and strokingto examine the effect of user interaction patterns on in-stability perception In the free exploration mode sub-jects were allowed to use the interaction pattern that

Figure 1 An illustration of the virtual textured surfaces and the two

coordinate frames used in our experiments Position of the stylus tip

was always measured in the world coordinate frame

Figure 2 An illustration of the parameters used in texture

rendering

398 PRESENCE VOLUME 13 NUMBER 4

they found most effective at discovering instability ofvirtual textures This mode was selected to be the mostchallenging interaction pattern for a haptic texture ren-dering system in terms of perceived instability In thestroking mode subjects were instructed to move thestylus laterally across the textured surface (ie along thex axis as shown in Figure 1) Stroking is the explorationmode most frequently employed by humans for textureperception and identification (Lederman amp Klatzky1987)

3 Psychophysical Experiments

In this section we describe the experimental de-sign and results of the psychophysical experiments con-ducted for a quantitative analysis of perceived instabilityduring haptic texture rendering The specific objectivesof these experiments were (1) to quantify the parameterspace that results in virtual textures that are perceived tobe stable and (2) to gather qualitative descriptions ofdifferent kinds of perceived instability

31 Experiment Design

311 Apparatus A PHANToM force-reflectinghaptic interface (model 10A SensAble TechnologiesWoburn MA) was used in all experiments to rendervirtual textured surfaces It was equipped with a stylus asan interaction tool and encoder gimbals for orientationsensing This PHANToM model has a nominal posi-tional resolution of 003 mm and a nominal maximumstiffness of 35 Nmm

312 Stimuli The virtual texture models used inthe experiments had two parameters amplitude A andwavelength L of the 1D sinusoidal gratings Three val-ues of A (05 10 and 20 mm) and three values of L(10 20 and 40 mm) were tested resulting in a totalof nine textured surface profiles These sinusoidal grat-ings can be well constructed due to the 003 mm nomi-nal positional resolution of the PHANToM Each sur-face profile was rendered with the two texture renderingmethods described in Equations 2 and 3 with stiffness

K as the parameter It follows that the three parametersA L and K along with the texture rendering methodsuniquely defined the stimuli used in this study

Due to the fact that the PHANToM workspaceboundary exhibits inferior dynamics performance thevirtual textured surface was restricted to a 15 cm 15cm region located near the center of the PHANToMworkspace

313 Subjects Three subjects participated inthese experiments One subject (S1 male) was an expe-rienced user of the PHANToM haptic interface Theother two subjects (S2 and S3 females) had not usedany haptic interface prior to this study The average ageof the subjects was 263 years old All subjects are right-handed and reported no known sensory or motor ab-normalities with their upper extremities

314 Conditions The independent variablesemployed in the experiments were texture renderingmethod exploration mode and amplitude and wave-length values of sinusoidal surface profiles Four experi-ments defined by the combinations of the two texturerendering methods and the two exploration modeswere conducted There were nine conditions (3 A 3L values) per experiment (see Table 1)

The dependent variable measured in all 36 experi-mental conditions (4 experiments 9 conditions perexperiment) was the maximum stiffness KT below whichthe rendered textured surface was perceived to be stable

315 Procedure All subjects went through ini-tial training to develop criteria for the perception of in-stability of a virtual textured surface During the train-ing the subject chose the texture rendering method andselected the values of A L and K The subject was in-formed that the virtual textures were rendered as 1Dsinusoidal gratings The subject was instructed to regardany sensation that felt unrealistic based on his or herexperience of real textures as an indication of perceivedinstability Each subject spent approximately one houron training

The method of limits (Gescheider 1985) was used inall experiments Given a pair of A and L values within

Choi and Tan 399

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

cally rendered textures are guaranteed to be perceptuallyldquocleanrdquo (ie free of artifacts)

In general there are two major sources of perceivedinstability during haptic texture rendering an improperenvironment model and unstable control of the hapticinterface Realistic texture rendering requires the envi-ronment model to faithfully follow the physical proper-ties of the objects being rendered In practice howeverthe environment model is usually an approximation tothe underlying physics with the goal of inducing a tar-get percept such as the roughness of a textured surfaceTo the extent that such simplification induces percep-tual artifacts the user can perceive a virtual surface to beunrealistic or unstable The stability of the haptic inter-face in a control sense is also a necessary condition forrealistic rendering Perception of buzzing and chatteringare likely caused by control instability Unbounded be-havior of a controller unmodeled dynamics of a hapticinterface quantization noise of encoders energy instill-ing effects of a zero-order-hold converter and asyn-chronous switch time can all lead to control instability(Gillespie amp Cutkosky 1996) Although many controltheoretical studies have tackled the stability problem ofvirtual walls using a simplified one-DoF interactionmodel (Gillespie amp Cutkosky 1996 Adams amp Han-naford 1999 Miller Colgate amp Freeman 2000 Han-naford Ryu amp Kim 2001 Hannaford amp Ryu 2002)these techniques cannot be easily extended to the inves-tigation of the stability of haptic texture rendering Do-ing so requires modeling haptic interaction as a multi-DoF problem and extending the surface model from asingle flat surface to complex curved surfaces therebysignificantly increasing the complexity of the analysis

Any study of perceived instability has to take into ac-count the effect of both environment modeling andcontrol stability In addition the human decision pro-cess also plays a role in the perceived quality of virtualtextures The experiments reported in this paper weredesigned to address the following two questions withhuman subjects

Under what conditions do human users perceiveinstability from virtual haptic textures rendered witha force feedback device

What are the proximal stimuli that are responsiblefor the perception of instability and why

In the first set of experiments we quantified the max-imum rendering stiffness under which virtual texturesare perceived to be realistic We conducted psychophysi-cal experiments using various texture model parameterstexture rendering methods and exploration modesQualitative descriptions of different kinds of instabilitywere also collected in these experiments In the secondset of experiments we identified the proximal stimulithat caused the perception of instability This was ac-complished through measurements and analyses of thesignals recorded on the stylus that interfaced the humanhand with the haptic interface Some of the results havebeen published earlier in Choi and Tan (2002a 2002b)

The remainder of this paper is organized as followsThe methods that are common to both psychophysicaland measurement experiments are laid out in section 2The specific methods and results for the psychophysicalexperiments are presented in section 3 Those for mea-surement experiments can be found in section 4 Weconclude the paper with a general discussion in section 5

2 Rendering and Exploration of VirtualTextures Used in Our Experiments

In this section we provide the details of the tex-ture model rendering methods and exploration modesused in both psychophysical and measurement experi-ments

21 Texture Model

The virtual textured surfaces were modeled as 1Dsinusoidal gratings superimposed on a flat surface Thisunderlying flat surface defined by z 0 in the worldcoordinate frame of the PHANToM formed a verticalwall facing the user of the PHANToM (see Figure 1)The sinusoidal grating was described by

z A sin ([2L]x) A

where A and L denote the amplitude and (spatial) wave-

Choi and Tan 397

length respectively (see Figure 2) Sinusoidal gratingshave been widely used as the basic building blocks fortextured surfaces for studies on haptic texture percep-tion (Lederman et al 1999 Weisenberger et al 2000)They have also been used as the basis of a function setfor modeling real haptic textures (Wall amp Harwin1999)

22 Texture Rendering Method

Two basic texture rendering methods were em-ployed in the current study Both methods use a spring

model to calculate the magnitude of the rendered forceas K d(t) where K is the stiffness of the textured sur-face and d(t) is the penetration depth of the stylus attime t (see Figure 2) The penetration depth is calcu-lated as follows

dt 0 if pzt 0A sin 2 pxtL A pzt if pzt 0 (1)

where p(t) (px(t) py(t) pz(t)) is the position of the tipof the stylus

The two methods differ in the way the force direc-tions are rendered The first method introduced byMassie (1996) renders a force Fmag(t) with a constantdirection normal to the underlying flat wall of the tex-tured surface The second method proposed by Ho etal (1999) renders a force Fvec(t) with varying directionssuch that it remains normal to the local microgeometryof the sinusoidal texture model Mathematically

Fmagt Kdt nW (2)

Fvect Kdt nT pt (3)

where nW is the normal vector of the underlying flatwall and nT(p(t)) is the normal vector of the texturedsurface at p(t) Both methods keep the force vectors inthe horizontal plane (zx plane in Figure 1) therebyeliminating the effect of gravity on rendered forces

The two texture rendering methods are natural exten-sions of virtual flat wall rendering techniques Perceptu-ally they are very different textures rendered by Fvec(t)feel rougher than those rendered by Fmag(t) with thesame texture model Textures rendered by Fvec(t) canalso feel sticky sometimes

23 Exploration Mode

An exploration mode refers to a stereotypical pat-tern of the motions that humans employ to perceive acertain object attribute through haptic interaction (Le-derman amp Klatzky 1987) In our experiments we testedtwo exploration modes free exploration and strokingto examine the effect of user interaction patterns on in-stability perception In the free exploration mode sub-jects were allowed to use the interaction pattern that

Figure 1 An illustration of the virtual textured surfaces and the two

coordinate frames used in our experiments Position of the stylus tip

was always measured in the world coordinate frame

Figure 2 An illustration of the parameters used in texture

rendering

398 PRESENCE VOLUME 13 NUMBER 4

they found most effective at discovering instability ofvirtual textures This mode was selected to be the mostchallenging interaction pattern for a haptic texture ren-dering system in terms of perceived instability In thestroking mode subjects were instructed to move thestylus laterally across the textured surface (ie along thex axis as shown in Figure 1) Stroking is the explorationmode most frequently employed by humans for textureperception and identification (Lederman amp Klatzky1987)

3 Psychophysical Experiments

In this section we describe the experimental de-sign and results of the psychophysical experiments con-ducted for a quantitative analysis of perceived instabilityduring haptic texture rendering The specific objectivesof these experiments were (1) to quantify the parameterspace that results in virtual textures that are perceived tobe stable and (2) to gather qualitative descriptions ofdifferent kinds of perceived instability

31 Experiment Design

311 Apparatus A PHANToM force-reflectinghaptic interface (model 10A SensAble TechnologiesWoburn MA) was used in all experiments to rendervirtual textured surfaces It was equipped with a stylus asan interaction tool and encoder gimbals for orientationsensing This PHANToM model has a nominal posi-tional resolution of 003 mm and a nominal maximumstiffness of 35 Nmm

312 Stimuli The virtual texture models used inthe experiments had two parameters amplitude A andwavelength L of the 1D sinusoidal gratings Three val-ues of A (05 10 and 20 mm) and three values of L(10 20 and 40 mm) were tested resulting in a totalof nine textured surface profiles These sinusoidal grat-ings can be well constructed due to the 003 mm nomi-nal positional resolution of the PHANToM Each sur-face profile was rendered with the two texture renderingmethods described in Equations 2 and 3 with stiffness

K as the parameter It follows that the three parametersA L and K along with the texture rendering methodsuniquely defined the stimuli used in this study

Due to the fact that the PHANToM workspaceboundary exhibits inferior dynamics performance thevirtual textured surface was restricted to a 15 cm 15cm region located near the center of the PHANToMworkspace

313 Subjects Three subjects participated inthese experiments One subject (S1 male) was an expe-rienced user of the PHANToM haptic interface Theother two subjects (S2 and S3 females) had not usedany haptic interface prior to this study The average ageof the subjects was 263 years old All subjects are right-handed and reported no known sensory or motor ab-normalities with their upper extremities

314 Conditions The independent variablesemployed in the experiments were texture renderingmethod exploration mode and amplitude and wave-length values of sinusoidal surface profiles Four experi-ments defined by the combinations of the two texturerendering methods and the two exploration modeswere conducted There were nine conditions (3 A 3L values) per experiment (see Table 1)

The dependent variable measured in all 36 experi-mental conditions (4 experiments 9 conditions perexperiment) was the maximum stiffness KT below whichthe rendered textured surface was perceived to be stable

315 Procedure All subjects went through ini-tial training to develop criteria for the perception of in-stability of a virtual textured surface During the train-ing the subject chose the texture rendering method andselected the values of A L and K The subject was in-formed that the virtual textures were rendered as 1Dsinusoidal gratings The subject was instructed to regardany sensation that felt unrealistic based on his or herexperience of real textures as an indication of perceivedinstability Each subject spent approximately one houron training

The method of limits (Gescheider 1985) was used inall experiments Given a pair of A and L values within

Choi and Tan 399

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

length respectively (see Figure 2) Sinusoidal gratingshave been widely used as the basic building blocks fortextured surfaces for studies on haptic texture percep-tion (Lederman et al 1999 Weisenberger et al 2000)They have also been used as the basis of a function setfor modeling real haptic textures (Wall amp Harwin1999)

22 Texture Rendering Method

Two basic texture rendering methods were em-ployed in the current study Both methods use a spring

model to calculate the magnitude of the rendered forceas K d(t) where K is the stiffness of the textured sur-face and d(t) is the penetration depth of the stylus attime t (see Figure 2) The penetration depth is calcu-lated as follows

dt 0 if pzt 0A sin 2 pxtL A pzt if pzt 0 (1)

where p(t) (px(t) py(t) pz(t)) is the position of the tipof the stylus

The two methods differ in the way the force direc-tions are rendered The first method introduced byMassie (1996) renders a force Fmag(t) with a constantdirection normal to the underlying flat wall of the tex-tured surface The second method proposed by Ho etal (1999) renders a force Fvec(t) with varying directionssuch that it remains normal to the local microgeometryof the sinusoidal texture model Mathematically

Fmagt Kdt nW (2)

Fvect Kdt nT pt (3)

where nW is the normal vector of the underlying flatwall and nT(p(t)) is the normal vector of the texturedsurface at p(t) Both methods keep the force vectors inthe horizontal plane (zx plane in Figure 1) therebyeliminating the effect of gravity on rendered forces

The two texture rendering methods are natural exten-sions of virtual flat wall rendering techniques Perceptu-ally they are very different textures rendered by Fvec(t)feel rougher than those rendered by Fmag(t) with thesame texture model Textures rendered by Fvec(t) canalso feel sticky sometimes

23 Exploration Mode

An exploration mode refers to a stereotypical pat-tern of the motions that humans employ to perceive acertain object attribute through haptic interaction (Le-derman amp Klatzky 1987) In our experiments we testedtwo exploration modes free exploration and strokingto examine the effect of user interaction patterns on in-stability perception In the free exploration mode sub-jects were allowed to use the interaction pattern that

Figure 1 An illustration of the virtual textured surfaces and the two

coordinate frames used in our experiments Position of the stylus tip

was always measured in the world coordinate frame

Figure 2 An illustration of the parameters used in texture

rendering

398 PRESENCE VOLUME 13 NUMBER 4

they found most effective at discovering instability ofvirtual textures This mode was selected to be the mostchallenging interaction pattern for a haptic texture ren-dering system in terms of perceived instability In thestroking mode subjects were instructed to move thestylus laterally across the textured surface (ie along thex axis as shown in Figure 1) Stroking is the explorationmode most frequently employed by humans for textureperception and identification (Lederman amp Klatzky1987)

3 Psychophysical Experiments

In this section we describe the experimental de-sign and results of the psychophysical experiments con-ducted for a quantitative analysis of perceived instabilityduring haptic texture rendering The specific objectivesof these experiments were (1) to quantify the parameterspace that results in virtual textures that are perceived tobe stable and (2) to gather qualitative descriptions ofdifferent kinds of perceived instability

31 Experiment Design

311 Apparatus A PHANToM force-reflectinghaptic interface (model 10A SensAble TechnologiesWoburn MA) was used in all experiments to rendervirtual textured surfaces It was equipped with a stylus asan interaction tool and encoder gimbals for orientationsensing This PHANToM model has a nominal posi-tional resolution of 003 mm and a nominal maximumstiffness of 35 Nmm

312 Stimuli The virtual texture models used inthe experiments had two parameters amplitude A andwavelength L of the 1D sinusoidal gratings Three val-ues of A (05 10 and 20 mm) and three values of L(10 20 and 40 mm) were tested resulting in a totalof nine textured surface profiles These sinusoidal grat-ings can be well constructed due to the 003 mm nomi-nal positional resolution of the PHANToM Each sur-face profile was rendered with the two texture renderingmethods described in Equations 2 and 3 with stiffness

K as the parameter It follows that the three parametersA L and K along with the texture rendering methodsuniquely defined the stimuli used in this study

Due to the fact that the PHANToM workspaceboundary exhibits inferior dynamics performance thevirtual textured surface was restricted to a 15 cm 15cm region located near the center of the PHANToMworkspace

313 Subjects Three subjects participated inthese experiments One subject (S1 male) was an expe-rienced user of the PHANToM haptic interface Theother two subjects (S2 and S3 females) had not usedany haptic interface prior to this study The average ageof the subjects was 263 years old All subjects are right-handed and reported no known sensory or motor ab-normalities with their upper extremities

314 Conditions The independent variablesemployed in the experiments were texture renderingmethod exploration mode and amplitude and wave-length values of sinusoidal surface profiles Four experi-ments defined by the combinations of the two texturerendering methods and the two exploration modeswere conducted There were nine conditions (3 A 3L values) per experiment (see Table 1)

The dependent variable measured in all 36 experi-mental conditions (4 experiments 9 conditions perexperiment) was the maximum stiffness KT below whichthe rendered textured surface was perceived to be stable

315 Procedure All subjects went through ini-tial training to develop criteria for the perception of in-stability of a virtual textured surface During the train-ing the subject chose the texture rendering method andselected the values of A L and K The subject was in-formed that the virtual textures were rendered as 1Dsinusoidal gratings The subject was instructed to regardany sensation that felt unrealistic based on his or herexperience of real textures as an indication of perceivedinstability Each subject spent approximately one houron training

The method of limits (Gescheider 1985) was used inall experiments Given a pair of A and L values within

Choi and Tan 399

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

they found most effective at discovering instability ofvirtual textures This mode was selected to be the mostchallenging interaction pattern for a haptic texture ren-dering system in terms of perceived instability In thestroking mode subjects were instructed to move thestylus laterally across the textured surface (ie along thex axis as shown in Figure 1) Stroking is the explorationmode most frequently employed by humans for textureperception and identification (Lederman amp Klatzky1987)

3 Psychophysical Experiments

In this section we describe the experimental de-sign and results of the psychophysical experiments con-ducted for a quantitative analysis of perceived instabilityduring haptic texture rendering The specific objectivesof these experiments were (1) to quantify the parameterspace that results in virtual textures that are perceived tobe stable and (2) to gather qualitative descriptions ofdifferent kinds of perceived instability

31 Experiment Design

311 Apparatus A PHANToM force-reflectinghaptic interface (model 10A SensAble TechnologiesWoburn MA) was used in all experiments to rendervirtual textured surfaces It was equipped with a stylus asan interaction tool and encoder gimbals for orientationsensing This PHANToM model has a nominal posi-tional resolution of 003 mm and a nominal maximumstiffness of 35 Nmm

312 Stimuli The virtual texture models used inthe experiments had two parameters amplitude A andwavelength L of the 1D sinusoidal gratings Three val-ues of A (05 10 and 20 mm) and three values of L(10 20 and 40 mm) were tested resulting in a totalof nine textured surface profiles These sinusoidal grat-ings can be well constructed due to the 003 mm nomi-nal positional resolution of the PHANToM Each sur-face profile was rendered with the two texture renderingmethods described in Equations 2 and 3 with stiffness

K as the parameter It follows that the three parametersA L and K along with the texture rendering methodsuniquely defined the stimuli used in this study

Due to the fact that the PHANToM workspaceboundary exhibits inferior dynamics performance thevirtual textured surface was restricted to a 15 cm 15cm region located near the center of the PHANToMworkspace

313 Subjects Three subjects participated inthese experiments One subject (S1 male) was an expe-rienced user of the PHANToM haptic interface Theother two subjects (S2 and S3 females) had not usedany haptic interface prior to this study The average ageof the subjects was 263 years old All subjects are right-handed and reported no known sensory or motor ab-normalities with their upper extremities

314 Conditions The independent variablesemployed in the experiments were texture renderingmethod exploration mode and amplitude and wave-length values of sinusoidal surface profiles Four experi-ments defined by the combinations of the two texturerendering methods and the two exploration modeswere conducted There were nine conditions (3 A 3L values) per experiment (see Table 1)

The dependent variable measured in all 36 experi-mental conditions (4 experiments 9 conditions perexperiment) was the maximum stiffness KT below whichthe rendered textured surface was perceived to be stable

315 Procedure All subjects went through ini-tial training to develop criteria for the perception of in-stability of a virtual textured surface During the train-ing the subject chose the texture rendering method andselected the values of A L and K The subject was in-formed that the virtual textures were rendered as 1Dsinusoidal gratings The subject was instructed to regardany sensation that felt unrealistic based on his or herexperience of real textures as an indication of perceivedinstability Each subject spent approximately one houron training

The method of limits (Gescheider 1985) was used inall experiments Given a pair of A and L values within

Choi and Tan 399

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

each of the four experiments a total of 100 series oftrials (50 ascending series and 50 descending series)were conducted Each ascending series started with astiffness value of Kmin 00 Nmm (ie no force) thatwas always perceived to be stable The subject wouldrespond ldquostablerdquo (by pressing a designated key on thekeyboard) The K value was then increased by K

002 Nmm The subject would feel the virtual tex-tured surface again and respond ldquostablerdquo or ldquounstablerdquoaccording to the perception As long as the subjectchose to report ldquostablerdquo the K value was incrementedby the same K amount for each subsequent trial Anascending series was terminated when the subject re-versed the response from ldquostablerdquo to ldquounstablerdquo Thevalue of K K2 was then recorded as the estimatedthreshold for this ascending series where K was thestiffness of the last trial with a ldquostablerdquo response

Each descending series started with a stiffness value ofKmax 06 Nmm This value was selected based onthe preliminary finding that no textured surface felt sta-ble at this K value The same step size of K 002Nmm was used to decrease K values in each subse-quent trial A descending series was terminated whenthe subject reversed the response from ldquounstablerdquo toldquostablerdquo The value of K K2 was then recorded asthe estimated threshold for this descending series whereK was the stiffness of the last trial with an ldquounstablerdquoresponse With these chosen values of Kmin Kmax andK each ascending-descending series could last up to31 trials

The experiments proceeded as follows Each subjectperformed all nine conditions (3 A values 3 L values)in Experiment I first followed by those in ExperimentsII III and IV The order of the nine conditions within

each experiment was randomized for each subject Foreach pair of A and L values the order of the 50 ascend-ing and 50 descending series was also randomized

During all experiments subjects wore headphoneswith white noise to block the auditory cues emanatingfrom the PHANToM No visual rendering of the tex-tured surface was provided Instead the computer mon-itor displayed only text information on the current seriesnumber

The following instructions were given to the subjectsduring all experiments They were asked to hold thestylus lightly and to hold it like a pen For the free ex-ploration mode (Experiments I and III) the subjectswere asked to detect any sensations indicating instabilityusing whatever interaction style they had chosen Forthe stroking mode (Experiments II and IV) the sub-jects were instructed to concentrate on the detection ofsensations indicating instability while they moved thestylus back and forth along the x direction across thetextured surface They were asked to maintain a con-stant stroking velocity to the best of their ability

Typically it took about an hour for a subject to finishone experimental condition Each subject finished twoor three experimental conditions per day It took a totalof approximately 36 hours for each subject to completethe 36 experimental conditions A 10 minute break wasenforced after a subject had completed the 100 ascend-ing-descending series associated with one experimentalcondition This was necessary in order to prevent a car-ryover effect (ie surfaces presented after a series of par-ticularly unstable conditions might have been judged asmore stable in a subsequent experiment) Subjects werealso allowed to take a break whenever it was needed

Table 1 Experimental Conditions for Psychophysical Experiments

Experiment Texture rendering method Exploration mode A (mm) L (mm)

I Fmag(t) Free exploration 5 10 20 10 20 40II Fmag(t) Stroking 5 10 20 10 20 40III Fvec(t) Free exploration 5 10 20 10 20 40IV Fvec(t) Stroking 5 10 20 10 20 40

400 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

316 Stability of Nontextured Flat Wall Af-ter the completion of the main experiments one subject(S1) was tested with a nontextured flat wall using thesame procedure as described above The result of thistest served as a baseline value for stiffness threshold KT

32 Results

As discussed earlier 50 ascending and 50 descend-ing series were conducted for each experimental condi-tion (ie each pair of A and L values within a main ex-periment) Figure 3 shows typical results for oneexperimental condition (subject S1 Fmag(t) strokingA 20 mm L 20 mm) The top panel shows thehistogram for all 50 ascending series the middle panelfor all 50 descending series and the bottom panel forcombined series The average of K values from the 50ascending series (026 Nmm) was greater than thatfrom the descending series (019 Nmm) This is typi-cal and reflects what is termed the ldquoerrors of habitua-tionrdquo (Gescheider 1985) It is a common practice tocompute the mean from the combined data (023Nmm) and regard it as an estimate of the stiffnessthreshold KT

Results from Experiment 1 (Fmag(t) free exploration)are shown in Figure 4 for the three subjects in separatepanels In each panel the stiffness thresholds are indi-cated by squares at their respective A and L values Themesh shows the fitted surface computed by linear re-gression analysis (see Equation 4) To help the readervisualize the spatial relationship between threshold datapoints (squares) and the fitted surface (mesh) straightlines are drawn between the centers of data points andthe corresponding points on the mesh with the same Aand L values The standard errors are not indicated inthe figures because they were very small (the averagestandard error was 0004 Nmm) The volume underthe mesh represents the parameter space within whichall virtual textured surfaces were perceived to be stableAs can be seen from Figure 4 these volumes were quitesmall for all subjects Subject S2 produced the largestvolume for stable texture rendering and S3 the smallestRecall that subject S1 was the only one who was experi-enced with the PHANToM device Therefore priorexperience with a force-reflective haptic interface didnot necessarily result in a particularly stringent or le-nient criterion for judging the stability of virtual tex-tured surfaces

A five-way ANOVA analysis (subject texture render-ing method exploration mode A and L) showed thatthere were significant differences among the three sub-jects tested [F(2 10791) 48457 p 0001] How-ever since all three plots in Figure 4 exhibited the samegeneral trends data from all subjects were pooled andsummarized in panel (a) of Figure 5 Also shown in Fig-ure 5 are the results from Experiments II III and IV inpanels (b) (c) and (d) respectively Overall the valuesof KT ranged from 00138 Nmm to 04527 Nmmfor all the conditions tested These values were quitesmall and the resulting textured surfaces felt very soft(like corduroy) They were also much smaller than thestiffness threshold measured with a nontextured wall(1005 0157 Nmm for subject S1) The effect ofexploration mode can be observed by comparing panel(a) with (b) and (c) with (d) The thresholds associatedwith the stroking mode [panels (b) and (d)] were largerthan those associated with the free-exploration mode[panels (a) and (c)] by an average difference of 0137

Figure 3 Typical histograms for one experimental condition

(Experiment II subject S1 A 20 mm L 20 mm) using the

method of limits

Choi and Tan 401

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

Nmm [F(1 10764) 598013 p 0001] Thethresholds for surfaces rendered with the Fmag(t)method [panels (a) and (b)] were statistically greaterthan those with the Fvec(t) method [panels (c) and (d)]by an average difference of 0099 Nmm [F(110764) 310344 p 0001]

The functional relationships between (A L) and KT

were estimated using the following form of a fittedequation

KT 0 A log2 A L log2 L

AL log2 A log2 L (4)

The estimated coefficients are listed in Table 2 for allexperiments They were computed by linear regressionanalysis for KT with two log-scaled continuous variables(log2 A and log2 L) and two categorical variables (tex-ture rendering method and exploration mode) as well astheir interaction terms (R2 05908 R2 is relativelysmall because we pooled the data of three subjectswhose results were significantly different) Note that thenonsignificant coefficients are set to zero in this table

The effects of the amplitude (A) and wavelength (L)of the sinusoidal gratings on the stiffness threshold KT

can be observed from Figure 5 and Table 2 In Experi-ments I and II it is evident from Figure 5 that KT de-creased as A increased The wavelength (L) had an ef-fect on KT only through the interaction term log2 A

log2 L but its effect was very small compared to that ofA (L 00 AL A) In Experiments III and IVincreasing A or L tended to result in lower or higherKT respectively unless KT was very small and theirinteraction was more apparent (AL of Experiments IIIand IV AL of Experiments I and II)

Subject debriefing revealed several types of perceivedinstability during haptic texture rendering In the freeexploration mode subjects reported that they perceivedthree types of apparent instabilities entry instabilityinside instability and ridge instability These terms re-flect the position of the stylus tip where the correspond-ing type of instability was perceived Entry instabilityrefers to the phenomenon that as the stylus approacheda point on the z 0 plane (see Figure 1) a high-frequency buzzing of the stylus sometimes occurred

Figure 4 Experimental results of Experiment I for all subjects The

stiffness thresholds KT are indicated by squares Regression surfaces

representing the boundary of KT for perceptually stable texture

rendering are also shown A solid line is drawn between the center of

a datum point and the corresponding point on the regression surface

with the same A and L values to help visualize the position of the

datum point

402 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

Inside instability was frequently associated with the ac-tion of poking It was most evident when the stylus waspushed deep into a virtual textured surface renderedwith Fvec(t) Ridge instability occurred when the styluswas pushed by the PHANToM into the valley of a sinu-soidal grating while the subject tried to maintain its po-sition on a ridge of the grating The subjects also re-

Table 2 Coefficients of Fitted Regression Equation

Experiment 0 A L AL

I 107 048 0 007II 281 156 0 007III 012 007 033 033IV 136 034 010 033

Figure 5 Results of psychophysical experiments The stiffness thresholds averaged over all three subjects are shown with open squares

Regression surfaces representing the boundary of KT for perceptually stable texture rendering are shown with meshes To help the reader

visualize the position of the data a solid line is drawn between the center of each datum point and the corresponding point on the regression

surface with the same A and L values

Choi and Tan 403

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

ported that entry and inside instabilities were morepronounced than ridge instability and therefore theymainly focused on the first two types of instability Inthe stroking mode the textured surface was perceivedto be unstable when the subject felt a buzzing noise inaddition to the vibrations resulting from the sinusoidalgratings being stroked across

33 Discussion

In the psychophysical experiments we measuredthe parameter space within which the subjects did notperceive any instability associated with the virtual tex-tures We tested two texture rendering methods usingsinusoidal gratings and two exploration modes Themaximum stiffness values under which no instability wasperceived turned out to be in the range of 00138 to04527 Nmm This range corresponded to surfacesthat were soft and spongy to the touch The stiffnessthreshold was much smaller than that of a flat surfacewith no texture (1005 Nmm)

Our finding that stroking resulted in a larger stiffnessthreshold than free exploration for the same renderingparameters is to be expected We recall that our subjectsrarely used stroking in the free exploration mode al-though it was allowed Instead they chose to positionthe stylus at various locations on or inside the virtualtextured surface to focus on the detection of any buzz-ing as an indication of instability Therefore in the freeexploration mode the subjects concentrated on the de-tection of vibrations in the absence of any other signalsIn the stroking mode the subjects always felt the vibra-tion due to the stylus stroking the virtual textured sur-face They had to detect additional noise in order todeclare the textured surface to be unstable Due to pos-sible masking of buzzing noise by the vibrations comingfrom the textured surface it is conceivable that subjectswere not able to detect instability with stroking as easilyas they could with static positioning of the stylus (freeexploration) In fact the subjects reported that the ex-periments with stroking were more difficult to performTherefore textured surfaces explored by stroking ap-peared to be more stable than those explored by pokingor static contact

Our finding that textures rendered with Fmag(t) re-sulted in a larger stiffness threshold than those renderedwith Fvec(t) is also consistent with the nature of thesetwo rendering methods While Fmag(t) imposed pertur-bations in the magnitudes of rendered forces onlyFvec(t) resulted in perturbations in both the directionsand the magnitudes of rendered forces The sometimesabrupt changes in force direction could cause virtualtextures rendered with Fvec(t) to be perceived as lessstable than those rendered with Fmag(t) To circumventthis problem Ho Basdogan and Srinivasan (1999)who originally proposed the Fvec(t) rendering methodhave developed a heuristic algorithm that interpolatesthe direction of a force vector between the normal tothe texture model and the normal to the underlyingsurface

To gain intuition into the effects of A or L on KT weconsider the derivative of the magnitude of the renderedforce Let g(t) Fmag(t) Fvec(t) and assume thatthe stylus is in contact with the textured surface FromEquations 1 to 3 we have

gt KA sin2

L pxt A pzt (5)

Differentiating g(t) with respect to the time variable tresults in

gt 2KAL cos2

L pxtpxt K pzt (6)

There are two terms in this equation that determinethe rate of change in the force magnitude The term onthe right K pz(t) responds to stylus motion in a direc-tion that is normal to the underlying plane [pz(t)] with again of K This is the same term that has been used informulating the virtual wall (with no texture) problemThe term on the left is due to textures on the virtualwall Here the lateral velocity of the stylus [px(t)] is am-plified with three constant gains (K A and 1L) andone variable gain that depends on the stylus position inthe lateral direction [px(t)] Increasing A or decreasingL results in a faster change in force magnitude whichcan cause a textured surface to be perceived as less stableor equivalently result in a smaller stiffness threshold KT

404 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

Of the three types of instability discovered by thesubjects the sensation associated with entry and insideinstability was that of buzzing and vibration The entryinstability was commonly observed for both texture ren-dering methods This instability may have resulted fromthe collision detection algorithm used in the experi-ments The collision detection algorithm declares a col-lision when the PHANToM stylus enters the underlyingplane Thus the penetration depth computed followingEquation 1 included step changes when the stylus en-tered and left the textured plane and these step changesmay have caused the perception of entry instability De-spite this known problem this collision detection algo-rithm is a basic and useful method that can be easilygeneralized to complex textured objects (see Choi ampTan 2003 for details)

The inside instability frequently observed in texturesrendered with Fvec(t) seems to be consistent with thenature of the texture rendering method When the sty-lus is positioned deep inside the texture surface Fvec(t)generates forces with relatively large magnitudes andfast direction changes This may have invoked the gen-eration of a high-frequency signal that the subjects de-scribed as buzzing

The sensation associated with ridge instability wasqualitatively different and was likely due to the inaccu-rate environment model of the textured surface When areal stylus rests on the ridge of a real surface with sinu-soidal gratings the reaction force and friction of thesurface combine to counterbalance the force exerted bythe userrsquos hand on the stylus thereby creating an equi-librium The force rendered by Fvec(t) however wasdetermined solely on the local texture geometry and didnot take into account the direction of user applied forceThis is illustrated in Figure 6 where it is assumed thatthe force applied by the user was normal to the planeunderneath the texture According to the environmentmodel the force applied by the PHANToM was alwaysin the direction of the surface normal nT(p(t)) As aresult the net force exerted on the tip of the stylus (thesum of the forces applied by the user and the PHAN-ToM) was directed towards the valley of the sinusoidalgrating Therefore the subject who tried to rest the sty-

lus on the ridge could feel the stylus being activelypushed into the valley

In summary the results of the psychophysical experi-ments showed that the parameter space for stable tex-ture rendering was too limited to be useful for virtualenvironment applications or psychophysical studies Asmentioned earlier the textured surfaces within the sta-ble rendering-parameter range felt like soft corduroyWe were not able to render harder or rougher texturedsurfaces without inducing the perception of instabilityIt was therefore necessary to investigate the characteris-tics and sources of signals that gave rise to perceivedinstability with the goal of eliminating them in order toincrease the useful parameter space for stable texturerendering

4 Measurement Experiment

In this section we report the results on the mea-surement of the proximal stimuli (position of the tip ofthe stylus force and acceleration) delivered to a sub-jectrsquos hand during the exploration of virtual texturesThe specific objectives of this experiment were (1) toisolate signals responsible for perceived instability (2) toidentify the signal components responsible for the per-ception of texture and instability respectively (3) toanalyze the intensity of the proximal stimuli in bothphysical and perceptual units and (4) to investigate thesources of signals causing the perception of instability

Figure 6 An illustration of the forces involved in ridge instability

Choi and Tan 405

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

41 Experiment Design

411 Apparatus The PHANToM force-reflecting device was used for both texture renderingand data collection The position of the tip of the stylusp(t) was measured using the position-sensing routinesin the GHOST library provided with the PHANToMThese routines read the optical encoders to sense jointangles of the PHANToM and converted them to a posi-tion of the stylus tip in the world coordinate frame

For force and acceleration measurement the PHAN-ToM was instrumented with two additional sensors Atriaxial forcetorque (FT) sensor (ATI Industrial Au-tomation Apex NC model Nano 17 with temperaturecompensation) was used to measure force delivered bythe PHANToM f(t) In order to minimize the struc-tural change to the PHANToM a new link with abuilt-in interface for the FT sensor was fabricated toreplace the last link (ie the link closest to the stylus) ofthe PHANToM (see Figure 7) The new link was of thesame length as the original one but weighed 60 g(13) more Force data were transformed into the styluscoordinate frame The origin of the stylus coordinateframe was always located at the tip of the stylus (iep(t)) and its z-axis coincides with the cylindrical axis ofthe stylus (see Figure 1)

Acceleration of the stylus was captured with a triaxialaccelerometer (Kistler Blairsville PA model 8794A500)The accelerometer was attached through a rigid mountthat was press-fitted to the stylus The attachment

added 118 g to the weight of the stylus Accelerationmeasurements a(t) were also taken in the stylus coordi-nate frame

The effects of the sensor attachments on the deviceperformance were investigated in terms of apparent in-ertia at the PHANToM stylus We measured the tip in-ertia of the original and instrumented PHANToMdevices along paths that passed the origin of the PHAN-ToM coordinate frame Two paths were chosen to beparallel to one of the axes of the PHANToM coordinateframe differing in the direction of tip movement duringthe measurements (a total of six paths) The results aresummarized in Table 3 It turned out that the tip inertiaalong the y-axis was affected most significantly by thethe addition of the two sensors In particular the appar-ent tip inertia in the y direction (the direction of grav-ity) was reduced by 1394 g This was due to the factthat the additional sensor weight increased the effectof gravity on the corresponding tip inertia The inertiaalong other directions changed in the range 138ndash294g and were therefore much less affected by the addi-tional sensor weight Since the forces used in our experi-ments for rendering textured surfaces were confined inthe x-z plane we concluded based on these measure-ments that the instrumented PHANToM was able toreproduce the stimuli that led to perceived instabilityduring the psychophysical experiments conducted ear-lier

412 Subjects Two subjects participated in themeasurement experiment (one male S1 and one fe-male S4) Their average age was 33 years old Both areright-handed and report no known sensory or motorabnormalities with their upper extremities Only S1 hadparticipated in the previous psychophysical experiments

Both subjects were experienced users of the PHAN-ToM device They were preferred over naive subjectsbecause our experiments required the subjects to placeor move the stylus in a particular manner in order tomaintain well-controlled conditions during data collec-tion

413 Experimental Conditions A total ofseven experimental conditions was employed (see Table

Figure 7 The PHANToM instrumented with a triaxial FT sensor

and an accelerometer

406 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

4) In the conditions using free exploration we col-lected data for the two primary instability categories ofentry and inside instability Note that inside instabilitywith Fmag(t) was not tested because this type of instabil-ity had not been observed during our previous psycho-physical experiments For stroking we recorded dataunder both stable and unstable conditions Both texturerendering methods were tested for the stroking mode

Whether a particular experimental condition resultedin the perception of instability depended on the valuesof the rendering parameters (A L and K) The valueslisted in Table 4 were selected based on the results ob-tained from our previous psychophysical experiments(see Figure 5) We chose stiffness values that were eitherone standard deviation below the measured stiffnessthresholds (for stable conditions) or one standard devia-tion above the thresholds (for unstable conditions)

414 Procedures For the experiments with thefree exploration mode the subjects were instructed tohold the stylus still near the textured surface (entry in-stability) or deep inside the textured surface (inside in-stability) They had to find a point in space where thesurface was clearly perceived to be unstable and main-tain that position Once the subject was satisfied withthe selected stylus position the experimenter initiateddata collection

For the experiments with the stroking mode the sub-jects were instructed to move the stylus laterally acrossthe virtual gratings They were required to maintain aconstant stroking speed to the best of their ability Afterthe subject had initiated stroking the experimenterstarted data collection

In all experimental conditions the subjects wereasked to hold the stylus like a pen (see Figure 7) Dur-

Table 3 Comparison of Apparent Tip Inertia of the Original and Instrumented PHANToM Devices

Tip movementdirection

Original PHANToMtip inertia (g)

Instrumented PHANToMtip inertia (g)

Effect of sensorattachment

x 2226 2088 138 g decreasex 2454 2240 214 g decreasey 2360 2788 428 g increasey 2824 1430 1394 g decreasez 972 1266 294 g increasez 1148 1010 138 g decrease

Table 4 Experimental Conditions for Measurement Experiment

Explorationmode

Texture renderingmethod

Perceptualcategory

Texture model parameters A(mm) L (mm) K (Nmm)

Free exploration Fmag(t) Entry instability 1 2 30Free exploration Fvec(t) Entry instability 1 2 30Free exploration Fvec(t) Inside instability 1 2 05Stroking Fmag(t) Stable 1 2 15Stroking Fmag(t) Unstable 1 2 40Stroking Fvec(t) Stable 1 2 15Stroking Fvec(t) Unstable 1 2 40

Choi and Tan 407

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

ing each trial 3D position force and acceleration datawere collected for 10 seconds at a sampling rate of1 kHz

415 Data Analysis Each 10-sec long time-domain signal was processed as follows Ten spectraldensities corresponding to the ten 1-sec segments of thesignal were computed and averaged for noise reductionWe used a flat-top window for the precise recovery ofthe magnitude of each spectral component (Smith1999) The frequency and magnitude of each promi-nent spectral component in its corresponding physicalunits were then calculated

In order to assess the perceived intensities of thesespectral peaks we compared their intensities in physicalunits to published human detection thresholds for sinu-soidal movements The human detection thresholds forvibrotactile stimuli depend on many factors includingbody site and contact area (Bolanowski Gesheider Ver-rillo amp Checkosky 1988) In our experiments the sty-lus was in contact with the distal pads of three fingers(thumb index finger and middle finger) and the webbetween the thumb and the index finger We thereforecompared our measurements to the detection thresh-olds taken at the distal pad of the middle finger (Ver-rillo 1971) and at the thenar eminence (Verrillo1963) For both sets of data we chose the thresholddata taken with contactor areas that are closest to ourexperimental setup (03 cm2 for finger tip and 13 cm2

for thenar eminence) It turned out that the thresholdcurves from these two body sites are quite similar attheir respectively chosen contact areas We thereforeused the detection thresholds for the thenar eminenceat 13 cm2 (Verrillo 1963) for our data analysis Theperceived magnitude of a given spectral peak from therecorded signals was computed as the difference be-tween the log of its intensity and the log of the humandetection threshold at the same frequency As is thecommon practice in psychophysics literature these per-ceived magnitudes are expressed in dB SL (sensationlevel)

For the stroking mode we estimated the location ofthe spectral peak corresponding to texture informationas follows Suppose that a subject explored the textured

wall by moving the stylus along the x-axis (see Figure 2)with a constant velocity of vx while maintaining contactwith the textured surface Then the magnitude of therendered force could be decomposed into two termsFrom Equations 2 and 3

Fmagt Fvect KA sin2vx

Lt KA pzt

(7)

assuming that px(0) 0 The left term delivered thetexture signal at frequency vxL and the right term pre-vented the penetration of the stylus into the texturedsurface Therefore ftex the estimated frequency for thespectral component responsible for texture perceptionwas

ftex xL (8)

where x was the average stroking velocity

42 Results

As an example of the collected data the experi-mental results for stable stroking using Fmag(t) areshown in Figure 8 for subject S4 Figures 8(a) 8(b)and 8(c) represent the position force and accelerationmeasurements respectively In each figure the mea-sured 3D time-domain data are plotted in the upperpanel and the corresponding power spectral densitiesare shown in the lower panel Note that the power spec-tral densities below 10 Hz are not shown because theyare likely to be 1f noises1 The predicted location ( ftex)of the spectral component for texture information com-puted using Equation 8 was 50 Hz for this data set Asexpected all spectral densities in Figure 8 showed spec-tral peaks around 50ndash60 Hz We therefore infer thatthe mechanical energy in this frequency band was in-deed responsible for the perception of the desired vir-tual texture It was also observed that no prominent

1 1f noise refers to a noise that starts at 0 Hz and rapidly decaysas frequency increases This noise is very commonly observed in mea-sured data (Smith 1999)

408 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

peaks appear at higher frequencies for this conditionFurthermore the locations of the distinct peaks mea-sured with all three sensors were highly consistent ex-cept for the high-frequency noise in the accelerometerdata caused by quantization error This consistencyacross sensor measurements turned out to be true for allexperimental conditions We therefore report only re-sults obtained from the position data pz(t) in the re-mainder of this paper Note that pz(t) showed the larg-est power spectral density among the three positionalvariables due to the fact that the normal vector of theunderlying flat wall was in the z direction (see Figure 1)

For the free exploration conditions where all render-ings were perceived to be unstable the power spectraldensities of pz(t) exhibited prominent spectral peaks inthe frequency range 192ndash240 Hz An example of suchdata (inside instability Fvec(t) subject S1) is shown inFigure 9(a) The upper panel contains the spectral den-sity function (solid line) along with the detectionthresholds taken from Verrillorsquos work (Verrillo 1963)(filled triangles) and the linearly-interpolated thresholdcurve (dotted line) The lower panel shows the differ-ence between the power spectral density and the detec-tion threshold curve The dotted line indicates the refer-ence line for 0 dB SL In both panels a vertical solidline is drawn to locate the peak in the spectral densityfunction We observe that only the signal componentsaround the spectral peak are significantly above the cor-responding absolute detection thresholds This fact wascommon to all experimental data for free exploration Itfollows that the energy in this high-frequency band (de-noted by fins) was responsible for the perception of in-stability

For the stroking data the predicted frequency ftex fortexture perception (see Equation 8) was used to locateits corresponding actual spectral peak (ftex) in the re-corded data The results in Table 5 show a close agree-ment between the values of ftex and ftex with an averageprediction error of 55 Hz Recall that ftex was estimatedunder the assumption that the subject moved the styluswith a constant stroking velocity This may have beenthe main source of discrepancy between the predictedand measured values of ftex

In the data measured under the conditions for per-

Figure 8 Experimental data for stable stroking (Fmag(t) subject S4) The

measured time-domain data are shown in the upper panels and their power

spectral densities in the lower panels The corresponding segments of Figure

8(b) (force) and 8(c) (acceleration) are indicated in Figure 8(a) (position)

Choi and Tan 409

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

ceptually stable stroking only one spectral componentthat delivered texture information appeared in thepower spectral densities An example is shown in Figure9(b) (Fmag(t) subject S4) Note that only one spectralpeak at 56 Hz (ie texture information) appears in thisfigure

For the conditions under which the subjects felt insta-bility during stroking usually two distinctive spectralcomponents were observed in the measured power spec-tral densities of pz(t) Figure 9(c) shows an example(Fmag(t) subject S1) of such cases This spectral densityfunction shows two prominent peaks at ftex 26 Hz(texture information) and fins 203 Hz (perception ofinstability)

Finally the perceived magnitudes of spectral compo-nents at ftex and fins are summarized in Table 6 for every

Figure 9 Average power spectral density of pz(t) and their

corresponding sensation levels The upper panels show the spectral

densities (solid lines) with the detection thresholds at the thenar

eminence (triangles and dashed lines) The lower panels show the

sensation levels as the difference between spectral densities and

detection thresholds The vertical lines mark the spectral components

for texture perception (ftex) and for perceived instability (fins)

Table 5 Predicted and Measured Locations of the SpectralPeaks for Texture Perception

Condition and subject ftex (Hz) ftex (Hz)

Stable stroking Fmag(t)S1 40 51S4 50 56

Unstable stroking Fmag(t)S1 29 26S4 39 41

Stable stroking Fvec(t)S1 37 40S4 57 62

Unstable stroking Fvec(t)S1 21 26S4 56 65

410 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

experimental condition and every subject The averagevalues of ftex and fins were 459 and 2137 Hz respec-tively The range of ftex (26ndash65 Hz) was well separatedfrom that of fins (192ndash240 Hz) Perceived magnitudesranged from 495 to 2361 dB SL for ftex componentsand 2121 to 4879 dB SL for fins components respec-tively

43 Discussion

In these experiments we measured position forceand acceleration signals experienced by the human handduring exploration of virtual textures Data were re-corded for free exploration and stroking modes usingtwo rendering methods under both stable and unstable

rendering conditions As we stated earlier our first goalwas to isolate signal components responsible for theperception of instability From our psychophysical ex-periments conducted earlier we learned that subjectsrelied on the detection of a ldquobuzzingrdquo vibration to de-clare a virtual textured surface to be unstable This indi-cated a signal with spectral components above 100 Hz(Tan 1996) Indeed we were able to isolate signalcomponents in the frequency range of 192ndash240 Hzfrom the measurements taken during unstable texturerendering conditions This happens to be the frequencyrange at which humans are most sensitive to vibrationalstimulation (Bolanowski et al 1988) From the mea-surement experiments we were also able to predictthen locate the spectral components responsible for tex-

Table 6 Intensities of Spectral Peaks (in Sensation Level) at Frequencies for TexturePerception (ftex) and Instability Perception (fins)

Condition and subject SL (dB) ftex (Hz) SL (dB) fins (Hz)

Entry instability Fmag(t)S1 mdash 3931 223S4 mdash 3348 221

Entry instability Fvec(t)S1 mdash 4784 238S4 mdash 3301 192

Inside instability Fvec(t)S1 mdash 4879 240S4 mdash 3092 205

Stable stroking Fmag(t)S1 495 51 mdashS4 891 56 mdash

Unstable stroking Fmag(t)S1 1567 26 2589 203S4 1098 41 2557 194

Stable stroking Fvec(t)S1 828 40 mdashS4 1383 62 mdash

Unstable stroking Fvec(t)S1 873 26 2121 208S4 2361 65 2653 203

Choi and Tan 411

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

ture information in the frequency range of 26ndash65 HzStimulation in this frequency range is usually perceivedto be ldquoroughrdquo and ldquoflutteringrdquo (Mountcastle TalbotDarian-Smith amp Kornhuber 1955) In order to charac-terize the perceptual intensities of measured signals themagnitudes of their spectral peaks were converted toperceived magnitudes in dB SL While the intensities ofthe signals conveying texture information were of inter-mediate magnitude (495ndash2361 dB SL) the high-frequency noises that gave rise to the perception of in-stability were generally ldquolouderrdquo (2121ndash4879 dB SL)in terms of perception2

The frequency ranges for texture and instability per-ception are not only well separated in their numericalvalues but in the neural mechanism mediating theirperception as well From an engineering point of viewone of the most important performance criteria for anystabilization technique for haptic rendering is the mini-mization of the loss of perceptual information due tostabilization (Hollerbach amp Johnson 2000) The factthat the two frequency ranges responsible for textureand instability perception are numerically well separatedsuggests that it ought to be possible to filter out thehigh-frequency spectral components responsible for in-stability perception without significantly altering thesignals containing texture information From a percep-tion point of view it is well established in haptic percep-tion literature that two mechanoreceptive afferent typesthe slowly adapting type 1 and the Pacinian system areresponsible for the perception of signals in the 26ndash65Hz and the 192ndash240 Hz frequency ranges respectively(Bolanowski et al 1988 Johnson Yoshioka amp Vega-Bermudez 2000) When signals in these two frequencyregions are combined they remain perceptually salientand distinctive Therefore our subjects were able tosimultaneously perceive the spectral peaks responsiblefor texture and instability

In an attempt to locate the sources of the high-frequency signals responsible for perceived instabilitywe measured a z-axis open-loop frequency response ofthe PHANToM with the tip of the stylus resting at the

origin of its world coordinate frame During the mea-surement the stylus was supported by a tight string an-chored from above which served to constrain the stylusto point towards the z-direction The stylus of thePHANToM could only move along the z-axis This fre-quency response is defined as

Hz f Pz f

Fz f (9)

where Fz( f ) is the Fourier transform of the z-axis forcecommand to the PHANToM and Pz( f ) is the Fouriertransform of the z-axis position of the stylus tip Theresult shown in Figure 10 indicates a mechanical reso-nance at 218 Hz A similar resonance frequency has alsobeen reported recently for model 15 of the PHANToM(Cavusoglu Feygin amp Tendick 2002) This resonanceis well within the 192ndash240 Hz frequency range There-fore it seems to be the source of the spectral peaks inthe frequency range that we had identified to be respon-sible for instability perception Since humans are mostsensitive to vibrations within this frequency range noiseproduced by this resonance tends to be ldquoloudrdquo percep-tually

5 General Discussion

We investigated the problem of perceived instabil-ity during haptic texture rendering with the aims toquantify conditions under which virtual textures are per-ceived to be stable to discover types of perceived insta-

2 According to Verrillo and Gescheider (1992) stimulation levelsexceeding 50ndash55 dB SL start to induce discomfort and fatigue

Figure 10 Frequency response of the PHANToM (model 10A)

measured at the origin and along the z-axis of its world coordinate

frame

412 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

bility and to understand the sources of perceived insta-bility Our work was performed in the context of apply-ing virtual textures to object surfaces rendered with aforce-reflecting haptic interface and of utilizing virtualtextures as stimuli in psychophysical experiments ontexture perception Conclusions drawn from the experi-ments reported here will be discussed with regard tothese two application areas

First we concluded that the parameter space for sta-ble haptic texture rendering using the PHANToM wastoo small to be useful in the sense that only texturesthat felt like soft corduroy could be rendered withoutany artifacts This result significantly restricts the typesof surface textures that can be properly rendered in avirtual environment or a psychophysical study Based onour results many published studies may have used vir-tual textures that contained perceived instability Towhat extent the results of these studies have beentainted by perceived instability is a question that needsto be carefully addressed

Second we found that stroking resulted in a largerstiffness threshold for stable texture rendering thanstatic contact or poking (free exploration) This result isin favor of psychophysical experiments where subjectshave to stroke the virtual surface in order to perceive itstexture Experimenters can explicitly prohibit subjectsfrom pushing too deeply into the textured surface inorder to preserve a stable rendering of the textures Invirtual environment applications however a user is usu-ally allowed to explore object surfaces freely The de-signer of virtual textures should therefore adhere to amore conservative parameter space for stable texturerendering

Third our measurements indicated that the frequencyof the ldquobuzzingrdquo noise that contributed to perceivedinstability was quite intense (2121ndash4879 dB sensationlevel) at a relatively high frequency (192ndash240 Hz) Inaddition measurements taken in our lab and othersrsquorevealed a mechanical resonance of the PHANToM de-vices at around 218 Hz These results along with thefact that texture rendering usually generates relativelyfast changing force commands stress the importance ofconsidering the high-frequency behavior of a force-

reflecting device such as quantization noise of encodersand flexibility of joints and links in both theoretical andexperimental studies on stable texture rendering

Fourth we were able to predict (from texture modeland user stroking velocity) and then locate (from ourposition force and acceleration measurements) the fre-quency components that conveyed texture informationduring the stroking mode This frequency range wasrelatively low (26ndash65 Hz) and was well separated fromthe frequency range contributing to instability percep-tion We therefore believe that it should be possible toremove the high-frequency noise from proximal stimuliwithout affecting the components conveying textureinformation Doing so will likely result in a significantincrease of the parameter space for stable haptic texturerendering

Finally the perceived instability frequently observedin our experiments resulted from two sources of per-ceived instability unstable control of the haptic inter-face and inaccurate environment dynamics Entry insta-bility and inside instabilities described as buzzing bythe subjects were due to the traditional control-relatedinstability (eg mechanical resonance) Ridge instabilityoccurred when force directions were varied based ononly texture geometry and stylus position without tak-ing into account friction or user-applied force Giventhe plethora of texture models that have been proposedthere is a pressing need for a better understanding ofhow these models and the associated rendering algo-rithms affect the perceived quality of haptic virtual tex-tures Our future work will continue to investigate theeffect of inaccurate environment dynamics on the per-ceived instability of a haptic texture rendering system

Acknowledgments

This work was supported in part by a National Science Foun-dation (NSF) Faculty Early Career Development (CAREER)Award under Grant 9984991-IIS and in part by an NSFaward under Grant 0098443-IIS The authors wish to thankthe anonymous reviewers for their insightful comments andsuggestions on an earlier version of this manuscript

Choi and Tan 413

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

References

Adams R J amp Hannaford B (1999) Stable haptic interac-tion with virtual environments IEEE Transactions on Robot-ics and Automation 15(3) 465ndash474

Bolanowski S J Jr Gesheider G A Verrillo R T ampCheckosky C M (1988) Four channels mediate the me-chanical aspects of touch Journal of Acoustical Society ofAmerica 84(5) 1680ndash1694

Cavusoglu M C Feygin D amp Tendick F (2002) A criti-cal study of the mechanical and electrical properties of thePHANToM haptic interface and improvements for highperformance control Presence Teleoperators and VirtualEnvironments 11(6) 555ndash568

Choi S amp Tan H Z (2002a) An analysis of perceptual in-stability during haptic texture rendering Proceedings of theTenth International Symposium on Haptic Interfaces for Vir-tual Environment and Teleoperator Systems 129ndash136

mdashmdashmdash (2002b) A study on the sources of perceptual insta-bility during haptic texture rendering Proceedings of theIEEE International Conference on Robotics and Automation1261ndash1268

mdashmdashmdash (2003) An experimental study of perceived instabilityduring haptic texture rendering Effects of collision detec-tion algorithm Proceedings of the Eleventh InternationalSymposium on Haptic Interfaces for Virtual Environmentand Teleoperator Systems 197ndash204

Costa M A amp Cutkosky M R (2000) Roughness percep-tion of haptically displayed fractal surfaces Proceedings of theASME Dynamic Systems and Control Division 69(2) 1073ndash1079

Fritz J P amp Barner K E (1996) Stochastic models forhaptic texture Proceedings of SPIErsquos International Sympo-sium on Intelligent Systems and Advanced ManufacturingmdashTelemanipulator and Telepresence Technologies III 34ndash44

Gescheider G A (1985) Psychophysics Method Theory andApplication (2nd ed) Mahwah NJ Erlbaum

Gillespie R B amp Cutkosky M R (1996) Stable user-specific haptic rendering of the virtual wall Proceedings ofthe ASME International Mechanical Engineering Congressand Exhibition 58 397ndash406

Hannaford B amp Ryu J-H (2002) Time-domain passivitycontrol of haptic interfaces IEEE Transactions on Roboticsand Automation 18(1) 1ndash10

Hannaford B Ryu J-H amp Kim Y S (2001) Stable con-trol of haptics In M McLaughlin (Ed) Touch in Virtual

Environments Proceedings of USC Workshop on Haptic In-terfaces (pp 47ndash70) Upper Saddle River NJ Prentice Hall

Ho C Basdogan C amp Srinivasan M A (1999) Efficientpoint-based rendering techniques for haptic display of vir-tual objects Presence Teleoperators and Virtual Environ-ments 8(5) 477ndash491

Hollerbach J M amp Johnson D E (in press) Virtual envi-ronment rendering Human and Machine Haptics

Hollins M amp Risner S R (2000) Evidence for the duplextheory of tactile texture perception Perception amp Psychophys-ics 62 695ndash705

Johnson K O amp Hsiao S S (1992) Neural mechanisms oftactual form and texture perception Annual Review of Neu-roscience 15 227ndash250

mdashmdashmdash (1994) Evaluation of the relative roles of slowly andrapidly adapting afferent fibers in roughness perception Cana-dian Journal of Physiology amp Pharmacology 72 488ndash497

Johnson K O Yoshioka T amp Vega-Bermudez F (2000)Tactile functions of mechanoreceptive afferents innervatingthe hand Journal of Clinical Neurophysiology 17(6) 539ndash558

Katz D (19251989) The world of touch Mahwah NJ Erl-baum

Kim L Kyrikou A Sukhatme G S amp Desbrun M(2002) An implicit-based haptic rendering technique Pro-ceedings of the IEEERSJ International Conference on Intel-ligent Robots and Systems 2943ndash2948

Klatzky R L amp Lederman S J (1999) Tactile roughnessperception with a rigid link interposed between skin andsurface Perception amp Psychophysics 61 591ndash607

LaMotte R H amp Srinivasan M A (1991) Surface micro-geometry Tactile perception and neural encoding In OFranzen amp J Westman (Eds) Information Processing in theSomatosensory Systems Wenner-Gren International Sympo-sium Series (pp 49ndash58) Houndmills UK MacMillan Press

Lederman S J amp Klatzky R L (1987) Hand movement Awindow into haptic object recognition Cognitive Psychology19 342ndash368

mdashmdashmdash (1999) Sensing and displaying spatially distributedfingertip forces in haptic interfaces for teleoperator and vir-tual environment systems Presence Teleoperators and Vir-tual Environments 8 86ndash103

Lederman S J Klatzky R L Hamilton C L amp RamsayG I (1999) Perceiving roughness via a rigid probe Psy-chophysical effects of exploration speed and mode of touchHaptics-e (httpwwwhaptics-eorg) 1(1)

Lederman S J amp Taylor M M (1972) Fingertip force

414 PRESENCE VOLUME 13 NUMBER 4

surface geometry and the perception of roughness by activetouch Perception amp Psychophysics 12 401ndash408

Massie T H (1996) Initial haptic explorations with thePHANToM Virtual touch through point interaction Un-published masterrsquos thesis Massachusetts Institute of Tech-nology Cambridge

Miller B E Colgate E amp Freeman R A (2000) Guaran-teed stability of haptic systems with nonlinear virtual envi-ronments IEEE Transactions on Robotics and Automation16(6) 712ndash719

Minsky M amp Lederman S J (1996) Simulated haptic tex-tures Roughness Proceedings of the ASME Dynamic Systemsand Control Division 58 421ndash426

Minsky M D R (1995) Computational haptics The sandpa-per system for synthesizing texture for a force-feedback displayUnpublished doctoral dissertation Massachusetts Instituteof Technology Cambridge

Mountcastle V B Talbot W H Darian-Smith I amp Korn-huber H H (1955) Neural basis of the sense of flutter-vibration Science 155 597ndash600

Okamura A M Dennerlein J T amp Howe R D (1998)Vibration feedback models for virtual environments Pro-ceedings of the IEEE International Conference on Roboticsand Automation 674ndash679

Siira J amp Pai D K (1996) Haptic texturingmdashA stochasticapproach Proceedings of the IEEE International Conferenceon Robotics and Automation 557ndash562

Smith S W (1999) The scientist and engineerrsquos guide to digi-tal signal processing (2nd ed) San Diego CA CaliforniaTechnical Publishing

Tan H Z (1996) Information transmission with a multi-finger tactual display Unpublished doctoral dissertationMassachusetts Institute of Technology Cambridge

Verrillo R T (1963) Effect of contactor area on the vibro-tactile threshold The Journal of the Acoustical Society ofAmerica 35(13) 1962ndash1966

mdashmdashmdash (1971) Vibrotactile thresholds measured at the fin-ger Perception amp Psychophysics 9(4) 329ndash339

Verrillo R T amp Gescheider G A (1992) Perception viathe sense of touch In I R Summers (Ed) Tactile aids forthe hearing impaired (pp 1ndash36) London Whurr PublishersLtd

Wall S A amp Harwin W S (1999) Modeling of surfaceidentifying characteristics using Fourier series Proceedings ofthe ASME Dynamic Systems and Control Division 67 65ndash71

mdashmdashmdash (2000) Effects of physical bandwidth on perceptionof virtual gratings Proceedings of the ASME Dynamic Systemsand Control Division 69 1033ndash1039

Weisenberger J M Krier M J amp Rinker M A (2000)Judging the orientation of sinusoidal and square-wave vir-tual gratings presented via 2-DOF and 3-DOF haptic inter-faces Haptics-e (httpwwwhaptics-eorg) 1(4)

Choi and Tan 415