Research Collection

Conference Paper

Towards Optimal Force Distribution for Walking Excavators

Author(s): Hutter, Marco; Leemann, Philipp; Stevsiic, Stefan; Michel, Andreas; Jud, Dominic; Figi, Ruedi; Caduff,Christian; Loher, Markus; Tagmann, Stefan; Hoepflinger, Mark; Siegwart, Roland

Publication Date: 2015

Permanent Link: https://doi.org/10.3929/ethz-a-010489032

Originally published in: http://doi.org/10.1109/ICAR.2015.7251471

Rights / License: In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.

ETH Library

Towards Optimal Force Distribution for Walking Excavators

Marco Hutter, Philipp Leemann, Stefan Stevsic, Andreas Michel, Dominic Jud, Mark Hoepflinger, Roland Siegwart1

Ruedi Figi, Christian Caduff, Markus Loher, and Stefan Tagmann2

Abstract— This paper presents the successful implementationof force control strategies on a 12 ton walking excavator tooptimize the ground reaction force distribution for betterstability, less terrain damage and to reduce the operationcomplexity. Using cleverly arrange standard industrial valvecomponents to separately control in- and out-flow of thehydraulic cylinders, we achieve accurate and fast joint torquecontrol purely based on pressure feedback. On the full systemlevel, we realize an automated force distribution to adjust thecenter of pressure and to level out the cabin of the machine.While the operator has still full control over the excavator, thisassistance system guarantees permanent ground contact andideal force distribution among the all four wheels independentof the level of the terrain. The proposed method significantlyimproves operability of the walking excavator in rough terrain.

I. INTRODUCTION

Walking excavators [1] belong to some of the most versa-tile but at the same time most complex mobile constructionsite machines. They are deployed in all kinds of specialscenarios that require advanced mobility such as in themountains, in rivers, or other hardly accessible areas. Incontrast to traditional tracked vehicles, the newest generationof multi-purpose excavators developed by Menzi Muck AG3

(e.g. M545 depicted in Fig. 1) is equipped with four legs thateach have three degrees of freedom and an actuated wheel.Thanks to this setup, these 12 ton construction site machinesare able to drive or walk over challenging terrain.

However, for all locomotion maneuvers, the human oper-ator must manually synchronize the motion of the four legs,meaning that he has to manually control the valve of each ofthe three cylinders per leg. This coordinatively challengingwork requires very skilled machinists, particularly whenmoving forward and excavating at the same time. Moreover,the permanent coordination necessary to move the machineover bumpy terrain can be very exhausting.

One of the main problems with four legs in simultaneousground contact is that a stiff and position controlled systemlike the M545 tends to be unstable in the sense that itcan be permanently wiggling over the diagonal axis ofcontact if the elevation of the terrain changes. Moreover,

*This research was supported by the Commission for Technology andInnovation (CTI) Switzerland through project 16004-1.

1 M. Hutter, P. Leemann, S. Stevsic, A. Michel, D. Jud, M. Hoepflinger,R. Siegwart are with the Autonomous Systems Lab, ETH Zurich, Switzer-land mahutter@ethz.ch

2 R. Figi, C. Caduff, M. Loher, S. Tagmann are with Menzi Muck AG,Switzerland rfi@menzimuck.com

3www.menzimuck.com

Fig. 1: Walking excavators from Menzi Muck such as the M545are deployed for work in challenging environments.

if the center of gravity (CoG) is close to the geometriccenter of the contact points, almost the entire weight of themachine is supported by one diagonal axis. This implieshigh peak forces which can damage the terrain or wearoff the suspension system of the excavator. The typicalapproach to overcome this deficiency in wheeled vehiclesis the use of passive suspension systems (e.g. [2]) thatcan lead to significantly improved climbing capabilities.Such mechanical intelligence is often applied in the field ofrover type systems. In hydraulics, a common concept is theapplication of a swing axle at the front or the rear suspension[1]. Unfortunately, all these passive suspension systems limitthe versatility of the machines.

The goal of this collaborative project between the Au-tonomous Systems Lab of ETH Zurich and Menzi MuckAG was to implement an active contact force adaptation on awalking excavator to path the road for autonomous operationrespectively locomotion of such vehicles in challenging envi-ronments. Enhancing the mobility of machines using activesuspension systems is applied in the walking communitysince the very first developments [3]. Most of the large scalelegged robots used for forestry work (e.g. Timberjack Hexa-pod Robot), to work in hazardous environments (e.g. MantisRobot [4] or TITAN XI [5]), or for demining (e.g. COMET-VI [6], [7]) use some sort of force, respectively impedancecontrol to adapt to the ground. Today, the most advanced fourlegged robots, such as the hydraulic quadrupeds BigDog [8]or HyQ [9], or the electrically actuated systems StarlETH[10] or MIT Cheetah [11] are fully force controlled. Thisenables these robots to actively regulate their (dynamic)behavior by precisely modulating the ground reaction forces.

Hydraulicdynamics

+ FL force

controller-

+ +

Friction

model+

-

Pressures + Friction Model Feedback

Fig. 2: Hydraulic force control: In the absence of force sensors, theoutput force (fl) must be estimated from the hydraulic force (fh).

Taking inspiration from the underlying control techniques,which are often based on some sort of virtual model control[12], the present paper outlines a method to implementactive contact force adaptation for a walking excavator usingexisting industrial components paired with advanced controlalgorithms. On an instrumented testbench, we first developeda force control strategy for the individual cylinders basedon pressure feedback. The characteristics and performanceof our approach using industrial components are comparedwith some of the most advanced hydraulic robots that areused in research [9] as well as with the highest performanthydraulic valves that could be found on market. Based on thislocal control approach, we then successfully implementedan overall force distribution algorithm to adjust the centerof pressure as well as to level out the operator cabin. Theproposed algorithms were tested on a full-scale Menzi MuckM545 respectively an A91 with the same chassis.

II. FORCE CONTROL OF A HYDRAULIC CYLINDER

Key element to an optimal ground contact force distri-bution is precise cylinder force control [13]. However, incontrast to some of the best hydraulic walking robots (e.g.[9]), economic reasons and system robustness inhibit the useof additional force sensors (Fig. 2). In fact, force control mustbe achieved solely based on pressure readings (fh) and hencerequires to estimate the load force fl respectively frictionforce ff based on state (piston position xp, piston velocityxp, and pressure P ) feedback. Furthermore, standard valvesthat are in use in construction site machines typically showvery different characteristics compared to walking robotslike HyQ [9], BigDog [8], or Atlas [14]. Hence, the firstbig challenge was to achieve accurate and fast-enough forcecontrol using standard components.

A. Test Bench Setup

To characterize the force control behavior and to tune con-trollers, we developed a single axis test bench (Fig. 3) con-sisting of an active cylinder with pressure sensors (RexrothRE 95138), a load cell to measure the output force (Mec-Sense PC4), a passive cylinder to produce the counter force,

position encoder force sensor pressuresensors

active cylinder passive cylinder (counter force)

Fig. 3: A single axis test setup equipped with a precise load cellwas used to characterize the performance and to tune the controller.

−100 −50 0 50 100

−5

0

5

v/v max (%)

f f/F

max(%

)

(a) Friction Measurements

−100 −50 0 50 100

−5

0

5

v/v max (%)

f f/F

max(%

)

(b) Stribeck Estimation

Fig. 4: Measurements (a) and estimation (b) of the friction proper-ties in a small cylinder of a walking robot (HyQ, red) and a largescale cylinder of a walking excavator (Menzi Muck M545, blue).

an incremental position sensor (Micro-Epsilon WPS-2100-MK77) to measure position and correspondingly flow, as wellas a standard proporional valve (Duplomatic BLS6).

B. Friction Identification and Modeling

To accurately control the mechanical force (fl) based onpressure readings in the chambers, a good friction model isrequired. For small scale cylinders which are typically usedin systems like HyQ (cross section area A = 2.01 ·10−4 m2,maximal hydraulic force Fmax = 3 kN), the friction prop-erties show a quite complex behavior dominated by staticand dynamic effects. As show in Fig. 4(a) (red), the frictionhas highly-nonlinear and dynamic properties. As shown in[15], the underlying effects can only be captured by arelatively complex LuGre model [16] respectively the modelsdeveloped by Tran et al. [17] which includes 23 parametersthat must be estimated.

In contrast thereto, the large scale cylinder employed inthe M545 (A=7.85·10−3 m2, Fmax=200 kN) shows muchsmaller relative friction with very limited dry friction ofless than 1% of the maximal force (blue). The correspond-ing characteristics can be accurately captured by a simpleStribeck curve [18] as depicted in Fig. 4(b). The frictionaleffects seem significantly less important than what has beenreported in other studies like [19], [20] and in contrast toother solutions implemented in excavators [21], accurateforce control seems possible without an additional load cell.

C. Valve Setup and Control

Today’s most advanced hydraulic robotic systems likeHyQ, BigDog, or Atlas are actuated by some of the highest-performance valves that can be found on market, e.g. MoogE24 respectively E30, which feature extremely high controlbandwidth (> 200 Hz) at very low weight. This is achievedby high-precision manufacturing such that the no-flow po-sition in the middle configuration of the valve (red box inFig. 5) reduces to a single point (critically closed valve).

Fig. 5: Double valve configuration to overcome the problem withlarge overlap.

In contrast thereto, standard valves of construction sitemachines often have a closed center spool with large over-laps. For example, the Duplomatic valve used in our testexcavator features an overlap of almost 50%. This is an easyand inexpensive way to minimize the flow through the valvewhen it is closed. However, when controlling the valve inpressure mode and hence around zero flow, this overlap canbe compensated but has to be crossed each time the measuredsignal crosses its reference. Beside energetic inefficiency, theoverlap leads to a deadband, which significantly decreasesthe controller performance.

To overcome this problem using standard components,the valve can be split in two independently controlled sec-tions which are connected to the same cylinder in parallel(Fig. 5, yellow background). Thereby, in- and out-flow ofboth chambers can be independently controlled and the over-respectively underlap of the valve setup (i.e. its ’closed’-position) can be electronically set such that the valves arecritically closed.

D. Performance Evaluation

We compared the performance of (A) the standard valve ofthe M545, (B) the standard valve in double configuration, and(C) a similar-sized valve with the highest performance wecould find on the market (Moog G671). For this comparison,only the hydraulic force control loop was considered (nofriction compensation as outlined in Sec. II-B). For pressureregulation, we implemented a simple PI controller:

uv = PI (∆p) . (1)

The external mobile pump of 4 kW electric power wasoperated at constant pressure of 200 bars, which correspondsto an average pressure level we expect in the excavator. Thehouses between the pump and valve as well as between thevalve and cylinders were selected to have about the lengthand diameter as in the M545.

The performance comparison in the step answer (Fig. 6)and the sine following (Fig. 7) indicates that the single valvehas already significant errors at 5 Hz and loses track entirelyat 10Hz . The performance estimation resulted in a bandwidthof about 5 Hz for the standard valve, about 15 Hz for the

t [s]0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

F [

N]

cyl

ind

er

× 104

0

1

2

3

4

refmoogsingledouble

Fig. 6: Step response comparison

t [s]1.45 1.5 1.55 1.6

F [N

]cyl

ind

er

× 104

-6

-4

-2

0

2

4

refmoogsingledouble

(a) 5Hz

t [s]0.65 0.7 0.75 0.8

× 104

-6

-4

-2

0

2

4

F [

N]

cyl

ind

er

refmoogsingledouble

(b) 10Hz

Fig. 7: Sine response comparison with frequencies of 5Hz (left)and 10Hz (right): While the single valve already shows significantdeviations at 5Hz, the double configuration can follow up to 15Hz.

double configuration and about 25 Hz for the Moog valve.Beside bandwidth, the double configuration convinces due toa good performance around zero flow.

III. PROTOTYPE IMPLEMENTATION ON M545

The control system for the prototype system is imple-mented on a National Instrument’s compactRIO (cRIO)system (Fig. 8). This real time processor and FPGA unitfeatures the hydraulic force control loops for each cylinderas well as the high-level force distribution and inclinationcontroller. The cRIO is connected to the system’s CANbus that allows to read user Joystick signals as well asjoint sensor measurements. Using the serial USB interface,a myAHRS+ IMU with integrated Kalman filter providesinformation about the cabin orientation (roll and pitch) andits acceleration. For cylinder control, we read in analogpressure signals of both chambers A and B and providecurrent command signal to the valve using standard DCmotor control units.

NI Compact RIO

Joystick (User Input)

CANJoint Sensors

Valve Controller

Dual Valve

P T

A B

UaIdes

IMU (chassis orientation)

USB

Cylinder

Pressure Sensor

U

Control500Hz

Fig. 8: The controller for the prototype system is implemented ona NI compactRIO system.

=const1

Linear regression: R² = 0.9967

0

10

20

30

40

0 10 20 30 40

Support Force F estimated form F [N]s cylinder

me

asu

red

Su

pp

ort

Fo

rce

F

[N

]s

Fig. 9: The estimated ground contact force from the cylinderpressure agrees very well with the ground truth measured byindustrial balance (R2 > 0.99, max error ≤ 1.5 kN)

IV. EXCAVATOR KINEMATICS

The kinematics of a single leg j ∈ [1...4] of the M545(Fig. 1) can be described by the 3 actuated joints ϕj anda set of link lengths Lj . With these paramters, the contactpoint can be generally expressed in a body fixed frame B as

Brj = Brj(ϕj ,Lj

)(2)

with the corresponding Jacobian

BJj =∂Brj

(ϕj ,Lj

)∂ϕj

. (3)

In the static case and for a given contact force BFj at thewheel, we can calculate the torques for the three joints usingJacobi-transposed mapping

τ j = BJTj BFj −

∑i

BJTj BGj . (4)

The second part corresponds to the gravitational contributionof all segment masses of the leg links (see [22], chapter5.1). Given the joint torques, a simple trigonometric mappingyields the individual cylinder forces.

In regular operation, the wheel has significantly less trac-tion force (tangential force) than normal force (IFj,z >>IFj,x) and since the roll angle of the chassis is relativelysmall, the torque for the three joints can be approximated by

τ j ≈ ∂Brj,z∂ϕj

T

Fj,s (5)

Fj,s = IFj,z − IG, (6)

with IG corresponding to the equivalent mass of the leg(including wheel) pressing on the ground. According to thedatasheets of Menzi Muck and [23], the supporting mass isabout 940 kg. In the following, we will refer to Fs as supportforce of the chassis. In case Fj,s becomes negative, leg j isnot supporting the chassis but is pulling it to the ground.

V. GROUND CONTACT FORCE CONTROL

This section investigates how accurately the ground con-tact force can be adjusted with the fully instrumented M545.

In a first experiment, we placed an industrial balanceunderneath the right hind leg while continuously increasingthe desired support force F dess from 0 to 35 kN. As depictedin Fig. 9, the ground reaction force estimated from pressure

10

15

20

25

angle ϕ1

[◦]

Support

Forc

e F

[kN

]s

−10 −5 0 5 10 15 20 25 30 35 40

80

100

Cyl

inder

Forc

eF

[kN

]cy

linder

des

meas

cylinder

I z

I z

F

F

F

Fig. 10: While changing the angle of the support cylinder (ϕ1 ∈[−7◦,+40◦], the ground contact force can be kept constant byadapting the cylinder force appropriately.

Fcylinder

1

Fig. 11: The left and right jack cylinders of the back legs (red) areforce controlled to vary the contact force distribution.

readings aligns very well with the balance measures (linearregression coefficient of R2 > 0.99). With a maximalmeasurement error of <1.3 kN, we achieve an accuracy betterthan 2%.

A second experiment was conducted to demonstrate thatthe ground contact force can be kept constant independentlyof the leg angle. To this end, the desired support force F dess

was set to 23 kN, the joint angle ϕ1 was manually movedfrom −7◦ to 40◦, and the actual ground contact forces weremeasured using the industrial balance. As shown in Fig. 10,the measured force from the balance (blue line) deviatesmaximally 0.7 kN from the constant desired ground contactforce (red line) while the cylinder force (green dashed) isappropriately changed as a function of the joint angle.

VI. FORCE DISTRIBUTION AND INCLINATION CONTROL

With four legs in simultaneous ground contact, aquadrupedal robot theoretically features six internal forces.This allows to change the ground contact force distributionwithout influencing the motion of the system [24]. Con-sidering only the most important one, the reaction forcein gravitational direction, there remains one internal degreeof freedom to modulate the distribution without influencingthe net vertical force as well as the moment around theroll and pitch rotational direction. Hence, the most simple,still versatile force control setup for the walking excavatorinvolves only the automated adjustment of the two jackcylinders of the back legs (red label in Fig. 11), whichdirectly produce the torque around the first flexion-extensionjoint (blue arrow). Active control of these two cylindersenables simultaneous control of the pitch angle (θ, yellow)and ground contact force distribution.

A. Virtual Model Pitch ControlA virtual force Fv which should be produced by the two

hind legs is calculated as a function of the inclination θ andthe gravity compensation force Fgc:

Fv = Kθ (θ0 − θ) + Fgc. (7)

Fgc is not only a function of the model of the system sincethe mass of the excavator can vary up to 50%, dependingon the load in the bucket and the unknown configurationof the upper machine part. Fortunately, since the motion ofthe system can be approximated as quasi-static, Fgc can bemeasured at runtime by adding up the forces at hind right(HR) and hind left (HL) legs. To filter out undesired dynamiceffects and sensory noise, this force is additionally lowpassfiltered (IIR, first order):

F igv = α(F iHR + F iHL

)+ (1− α)F i−1

gv . (8)

B. Force DistributionIn a second step, the virtual force Fv is distributed over

the hind legs according to the force distribution measured atthe front right (FR) and front left (FL) legs

FHR =FFR

FFR + FFLFv = ηFRFv (9)

FHL =FFL

FFR + FFLFv = ηFLFv. (10)

This creates a virtual contact point or center of pressure(CoP) positioned between the two hind legs such that themachine is virtually supported by a triangle. The location ofthe CoP is determined by the base line b and the distributionfactors η: At an equal distribution (which corresponds to amechanical swing axle), the CoP is in the middle (Fig. 12,black); if the boom and hence the CoG is rotated to theleft respectively right of the machine, the CoP is shifted tothe left (red) respectively right (blue) of the machine. Thismechanism effectively shifts the tilting axle of the supporttriangle such that stability is ensured at all time.

C. Operator ControlThe operator input at the front legs can be used for feed

forward compensation to improve the dynamic response:

FFFHR = FFFuFR (11)FFFHL = FFFuFL, (12)

with FFF being a tuning factor to achieve a smooth behaviorand uFR respectively uFL ∈ [−1, 1] the valve opening signalfor the front jack cylinder given at the operator joystick.

VII. EXPERIMENTS

The following set of experiments which are consideredas the most relevant working tasks in daily operation wereconducted with the identical control setup and parameters.To get a good performance evaluation of the active balancingmechanism, all tasks were additionally executed by the mostexperienced driver at Menzi Muck AG. They are documentedwith a movie1.

1http://youtu.be/fKRjXU6sA2M

kg

kg

kg

b

bFR

bFL

Fig. 12: The virtual contact point respectively CoP and henceground reaction force distribution (arrows next to the wheels) isadapted as a function of the measured CoG location.

A. Changing Level

Changing level of the machine requires to move all fourjack cylinders in the front and back. During manual control,the operator needs to open the corresponding valves usingtwo thumb joysticks per hand (Fig. 13(c), blue). Hence, thetwo front and two rear joints are usually sequentially moved(Fig. 13(a), blue). By enabling the proposed assistive forcedistribution, the level of the machine is determined only bythe two angles of the front legs and hence requires to moveonly the thumb joystick of the front legs (Fig. 13(c), red)while the machine automatically keeps its pitch orientationFig. 13(b).

B. Turning the Operator’s Cabin

Moving loads from one side of the machine to the otherimplies a substantial shift of the overall CoG and henceentails ground contact force redistribution. In case the CoPdefined by the desired force distribution among the hind legswould be kept constant, the CoG eventually falls outside thesupport triangle and the machine tips over. This is avoided byshifting the CoP depending on the force measurements at thefront axis. As illustrated in Fig. 14(a), the force distribution atthe hind legs (red solid) can be kept equal to the front legs(red dashed) independently of the cabin orientation thanksto an accurate force tracking in the two support cylinder(Fig. 14(b)).

C. Moving across Obstacles

The reaction of the excavator is tested when driving overan obstacle of 320 mm height with the left hind wheel(Fig. 15(a)). The experiment is split in three parts: First,the excavator is driven on the obstacle with the controllerdeactivated and no input from the operator. Second, theoperator tries to negotiate the obstacle as good as possible.

t [s]0 1 2 3 4 5

Join

t angle

s [°

]

-10-505

10

15

2025

F R2,manual

HR 1,manual

F R2,controller

HR 1,controller

(a) Joint angles of the right front and hind leg

t [s]0 1 2 3 4 5

Incl

ina

tion

[°]

-8

-6

-4

-2

0

θmanualθ

controllerθ

0

(b) Inclination of the chassis

t [s]0 1 2 3 4 5

input [-

1,1

]

-1

0u

F,manualu

H,manualu

F,controller

(c) Joystick inputs

Fig. 13: Lifting the body with manual control (blue) requirescoordination of four thumb joysticks (c) and can lead to largeinclination errors (b) while the automated controller (red) keepsthe inclination almost constant with only 2 user inputs.

turn angle [°]-80 -60 -40 -20 0 20 40 60 80

ratio

[0

,1]

0

0.2

0.4

0.6

0.8

1

ηHR, controller

ηFR, controller

ηHR,manualη

FR,manual

(a) Force ration

t [s]0 5 10 15 20 25 30 35

F [

N]

cylin

der

× 105

0.5

1

1.5

2F

HR,actF

HR,refF

HL,actF

HL,ref

(b) Force tracking

Fig. 14: In contrast to manual control (blue), the automated forcedistribution (red) of the hind (solid) and front (dashed) legs is almostequal while rotating the boom from the right (−90◦) to the left(90◦) (a). This is only possible due to accurate force tracking (b).

(a) Obstacle (b) Ramp

Fig. 15: Experimental setup for (a) driving over an obstacle of380 mm height and (b) up and down a ramp.

t [s]0 2 4 6 8 10 12 14 16 18

ratio

[-]

-0.2

0

0.4

0.2

0.6

ηno

ηmanual

ηcontroller

ηdesired

no chassis support

(a) Force distribution ratio ηHR at the right hind leg

t [s]0 2 4 6 8 10 12 14 16 18

Incl

ina

tion

[°]

-1.5

-1

-0.5

0

0.5

1

1.5θ

noθ

manualθ

controllerθ

desired

(b) Body inclination

t [s]0 2 4 6 8 10 12 14 16 18

Join

t angle

s [°]

2

6

18

22

14

10 HL 1,no

HL 1,manual

HL 1,controller

(c) Joint angle ϕ1

Fig. 16: The force distribution at hind legs (a) can be kept constantη = 0.5 and the inclination of the cabin (b) does barely changewhile driving across obstacle with the activated controller.

And third, the controller is activated which distributes theforces over the hind legs equally. As shown in Fig. 16(a),the support force distribution ηHR at the right hind legdrops below zero in the uncontrolled (yellow) and manuallycontrolled (blue) case, implying that the leg carries no massof the chassis (FHR,s < 0). With the activated controller(red), the cabin remains leveled (Fig. 16(b)), the joint angleϕHL1 is adapted (Fig. 16(c)), and the force ratio η = 0.5stays constant (Fig. 16(a)).

D. Moving on Uneven Ground

Finally, the excavator was tested when partially movingon an inclination (Fig. 15(b)). This is a typical workingsituation which involves manual adjustment of the front legsand automated adaptation of the hind legs.

As illustrated in Fig. 17, even a very experienced driveris not capable to achieve an acceptable distribution of theweight on all four wheels when driving manually (blue): Theforce distribution entirely shifts from the left to the right.

With activated contact force adaptation, the supportingforces remain evenly distributed. Moreover, the driver onlyneeds to control the front legs to adjust the roll angle or bodyheight. Tracking errors in the force and hence change in theforce distribution (e.g. at 10 s) occur only during movementof the front and hind legs which leads to flow saturation.

t [s]0 5 10 15 20 25 30 35 40

ratio

-0.4-0.2

00.20.40.60.8

11.21.4

ηmanual

ηcontroller

no chassis support (left)

no chassis support (right)

Fig. 17: Force distribution ratio of the hind cylinders when drivingup and down the ramp.

VIII. CONCLUSION

In this paper we presented a clever force control methodfor the two jack cylinders of the hind legs of a walkingexcavator. The proposed method accurately levels out thecabin and optimally distributes the contact force among thesupporting wheels to ensure stability and to avoid unneces-sary peak contact and cylinder forces. This assistance systemcan be activated or disabled upon the operators request anddoes not reduce versatility of the system in any case sincethe operator always keeps full control over the machine.

While a thorough user end-study could not be conducted inthe context of this study, the proposed assistance system hasbeen very well received by all users. It is very intuitive andneeds no additional explanation to be used. The outstandingbenefit was the fact that the active controller allows tomove the excavator across obstacles while being constantlystabilized on four legs. Thereby, the operator does not haveto worry about appropriate adjustment of the support legsbut can fully concentrate on the excavating task.

An important outcome for coming research is the accurateforce controllability that can be achieved with standardconstruction machine valve components and purely basedon pressure sensing. We could show that frictional effectscontribute only to about 1% of the maximal cylinder force.With our system, we can control the ground contact force toan accuracy of less than 150kg, which corresponds to 1.25%of the empty machine weight. With this tool at hand, our nextsteps will involve simultaneous control of all chassis cylinderand hence potentially full automation of the machine.

IX. ACKNOWLEDGEMENT

The authors would like to thank Moog Inc. for providinga modified G671 valve to evaluate the force control perfor-mance of our setup. Furthermore, we thank Jonas Buchlifor giving access to the hydraulic pump facility and ThiagoBoaventura for his technical support.

REFERENCES

[1] Horst Konig, Maschinen im Baubetrieb, Springer FachmedienWiesbaden, 2014.

[2] R Y Siegwart, P Lamon, T Estier, M Lauria, and R Piguet, “InnovativeDesign for Wheeled Locomotion in Rough Terrain,” Robotics andAutonomous Systems, vol. 40, no. 2-3, pp. 151–162, 2002.

[3] Shin-Min Song and Kenneth J. Waldron, “Machines That Walk: TheAdaptive Suspension Vehicle,” Nov. 1988.

[4] Matt Danton, “Mantis Robot,” 2015.

[5] R Hodoshima, Y Fukuda, T Doi, S Hirose, T Okamoto, and J Mori,“Development of TITAN XI: a quadruped walking robot to work onslopes,” in IEEE/RSJ International Conference on Intelligent Robotsand Systems (IROS), 2004, pp. 792–797.

[6] Addie Irawan and Kenzo Nonami, “Optimal impedance control basedon body inertia for a hydraulically driven hexapod robot walking onuneven and extremely soft terrain,” Journal of Field Robotics, vol. 28,no. 5, pp. 690–713, Sept. 2011.

[7] Addie IRAWAN, Kenzo NONAMI, Hiroshi OHROKU, YasunagaAKUTSU, and Shota IMAMURA, “Adaptive Impedance Control withCompliant Body Balance for Hydraulically Driven Hexapod Robot,”Journal of System Design and Dynamics, vol. 5, no. 5, pp. 893–908,2011.

[8] M Raibert, K Blankespoor, G Nelson, and R Playter, “BigDog, therough-terrain quadruped robot,” in Proceedings of the 17th WorldCongress, 2008, pp. 10823–10825.

[9] C. Semini, N. G. Tsagarakis, E. Guglielmino, M. Focchi, F. Cannella,and D. G. Caldwell, “Design of HyQ – a hydraulically and electricallyactuated quadruped robot,” Proceedings of the Institution of Mechan-ical Engineers, Part I: Journal of Systems and Control Engineering,vol. 225, no. 6, pp. 831–849, 2011.

[10] Marco Hutter, Christian Gehring, Mark A. Hopflinger, Michael Blosch,and Roland Siegwart, “Toward Combining Speed, Efficiency, Versatil-ity, and Robustness in an Autonomous Quadruped,” IEEE Transactionson Robotics, vol. 30, no. 6, pp. 1427–1440, Dec. 2014.

[11] Sangok Seok, Albert Wang, David Otten, Jeffrey Lang, and Sang-bae Kim, “Design principles for highly efficient quadrupeds andimplementation on the MIT Cheetah robot,” in IEEE InternationalConference on Robotics and Automation (ICRA), 2013, pp. 3307–3312.

[12] J. Pratt, P. Dilworth, and G. Pratt, “Virtual model control of a bipedalwalking robot,” International Conference on Robotics and Automation(ICRA), pp. 193–198, 1997.

[13] Thiago Cuhna Boaventura, Hydraulic Compliance Control of theQuadruped Robot HyQ, Ph.D. thesis, University of Genoa, Italy andIstituto Italiano di Tecnologia (IIT), 2013.

[14] Gabe Nelson, Aaron Saunders, Neil Neville, Ben Swilling, JoeBondaryk, Devin Billings, Chris Lee, Robert Playter, and Marc Raib-ert, “PETMAN: A Humanoid Robot for Testing Chemical ProtectiveClothing,” Journal of the Robotics Society of Japan, vol. 30, no. 4,pp. 372–377, 2012.

[15] Stefan Stevsic, Pressure-based Hydraulic Impedance Control, Masterthesis, ETH Zurich, 2014.

[16] Carlos Canudas de Wit, H Olsson, KJ Astrom, and P Lischinsky, “Anew model for control of systems with friction,” IEEE Transactionson Automatic Control, vol. 40, no. 3, pp. 419–425, 1995.

[17] Xuan Bo Tran, Nur Hafizah, and Hideki Yanada, “Modeling ofdynamic friction behaviors of hydraulic cylinders,” Mechatronics, vol.22, no. 1, pp. 65–75, 2012.

[18] Brian Armstrong-Helouvry, Pierre Dupont, and Carlos Canudas DeWit, “A survey of models, analysis tools and compensation methodsfor the control of machines with friction,” Automatica, vol. 30, no. 7,pp. 1083–1138, 1994.

[19] Andrew Alleyne and Rui Liu, “A simplified approach to force controlfor electro-hydraulic systems,” Control Engineering Practice, vol. 8,no. 12, pp. 1347–1356, Dec. 2000.

[20] Shaharam Tafazoli, Identification of Frictional E ects and StructuralDynamics for Improved Control of Hydraulic Manipulators, Ph.D.thesis, University of British Columbia, 1997.

[21] S. Tafazoli, S.E. Salcudean, K. Hashtrudi-Zaad, and P.D. Lawrence,“Impedance control of a teleoperated excavator,” IEEE Transactionson Control Systems Technology, vol. 10, no. 3, pp. 355–367, May2002.

[22] M Hutter, StarlETH & Co - Design and Control of Legged Robotswith Compliant Actuation, Ph.D. thesis, ETH Zurich, 2013.

[23] Andreas Michel, Simulation of a walking excavator, Master thesis,ETH Zurich, 2013.

[24] M Hutter, H Sommer, C Gehring, M. Hoepflinger, M Bloesch, andR Siegwart, “Quadrupedal locomotion using hierarchical operationalspace control,” The International Journal of Robotics Research (IJRR),vol. 33, no. 8, pp. 1062–1077, May 2014.