MORIA - A Robot with FuzzyControlled Behaviour

HartmutSurmannandLiliane Peters

GermanNationalResearchCenterfor InformationTechnology,Institutefor SystemDesignTechnology

53754SanktAugustin,GermanyE-mail: surmann@gmd.de.

Abstract. 1 MORIA is a low cost fuzzy controlledautonomousservicerobot.Themajorgoalof our experimentwasto developa human-like decisionstrategythat supportsautonomousnavigation of the systemin office buildings.This pa-per presentsthe variousstepsduring the designcycle, from specificationto itsimplementation.Experimentalresultsshow that the developedprototypereactsautonomouslyin anappropriatemannerto thevariousdynamicchangesencoun-teredin thetestenvironment.

1 Intr oduction

A new breadof robotsis enteringour daily life. They cleanthe floors at airports[1],carrysuit-casesin hotels[2] andguideyou within a museum[3]. Theserobots,knownasservicerobots,have four commonfeatures.(1) They work within structured envi-ronmentsthatare a-priori known. Officeenvironmentscanbeconsideredstructuredasthey have fixed building elementslike: corridors,halls, elevators,rooms,etc. (2) Thecomplexity of theenvironmentis high.Not only dothestructuredbuilding elementsdif-fer from building to building, but alsothenumberof elementswithin abuilding is quitehigh especiallyif we considerthat office building canhave from two to thirty floors.(3)Theenvironmentis highlydynamic.Theservicerobotshaveto operatetogetherwithhumansin thesameenvironment.Thatmeanssharingthesamepathandaccesspointsfrom oneclosedenvironmentalsection- corridor - to thenext one.Thecomplexity isincreasedeven moreif the servicerobotsoperatein groups.But not only mobile ob-jectsreflectthe dynamicchangesin the environment,staticobjectslike, desks,chair,boxes,etc.canblock or partially occludepartsof known paths.Their appearenceanddisappearancein theenvironmentcannot beforeseen(planned)a-priori. (4) Theusersare technicallyunskilledpersonnel.Therefore,notonly doesthehumaninterfaceto themachinehave to besimple,but anadditionalsourceof uncertaintyhasto betakenintoconsideration.Thesefour featuresrequirea high navigationautonomyof the robot intheoffice environment,anda strategy for efficient tasksolving - from collision avoid-anceto reschedulingof missions.Althoughthefirst stepshavebeenmade,thereis still

1 in Fuzzy Logic Techniquesfor Autonomous Vehicle Navigation (D. Driankov & A.Saffiotti (Eds.)), Series: Studies in Fuzzinessand Soft Computing. Vol. 61, SpringerVerlag Berlin, 2001, pp. 343 - 366 ISBN 3-7908-1341-9,http://www.springer.de/cgi-bin/searchbook.pl?isbn=3-7908-1341-9

a very long way to go until populationsof robots“li ve” within our buildings or evenapartments,andshareall heavy, unhealthy, or time-consummingchores,so improvingthe quality of our daily life. Until now, cooperationhasbeenmainly studiedbetweenonerobot anda humanor a few robotsin a room like environment,e.g. [4] [5], [6].Whatis still missingaretheexperimentswith largenumbersof robots(over50)drivingthroughbusybuildings.Thefirst steptowardthis vision is to developa navigationandplanningstrategy for a singlerobot thatcancopewith a service-like environment.Theincreasein complexity throughtheintroductionof a groupis thenext step.

Fuzzylogic theoryoffersa naturalway to integratetheserequirementsin onesys-tem.It hasalreadybeenshownthatfuzzyrulebasedsystems(FRBS)areimportanttoolsfor modellingcomplex systemswhere,dueto thecomplexity or theimprecisionof theprocessto becontrolledtheknowledgebaseis acquiredfrom humanexpertor from areferentialdatasetwith neuralor geneticalgorithms[7]. Theadvantageof fuzzy con-trol is obviouswhenthecontrolprocessis non-linear, hasa time dependentbehaviourandthe availablemeasurementshave a poor quality or a high uncertainty. The fuzzylogic theoryreplacescrisp valueslike “true” and“f alse”with a rangeof truth values,crisp measurementslike 0.5m or 1.2m,and linguistic termslike “near” or “default”.This linguistic descriptionloosenstherequiredaccuracy of thesensorialmeasurement.Thustheoutputof thesensorshaveto guaranteearangeof valuescorrespondingto pre-definedlinguistic termse.g.“near” insteadof absoluteerrordeviation.Thefuzzy termsin turnaredescribedby acontinuousmembershipfunctiondefinedoverthesamerangeof values.This descriptionmirrorsnaturallanguageconceptsandmakesit possibletoprocesslinguistic conceptslike the“distanceis short” in a precisecontrolsystem.Forcompletedescriptionof thefuzzy logic controlmethodsee[8].

In this chapterwe proposea recurrentfuzzy systemto implementtheautonomousbehaviour of the robot. It hasto be mentionedthat it is not the only approachusingfuzzy logic for robotnavigation.For comparisonsee[9–14],eachoneapplyingvariousnavigationstrategies.Theoriginality of ourapproachis theintroductionof aninferren-tial reasoningprocessthatsolvesinherentambiguitieswithin theenvironment.Thusthebehaviour of therobot is not only dependenton thesensorialinput but alsoon its pastbehaviour. This approachgivesahigh degreeof autonomyto therobot.Fromthepointof view of therobottheapproachcabedescribedto supportanhierarchicalarchitecturebasedon a high level planneranda low level executer[15]. Both levelsarebasedon afuzzy logic conceptandcopesuccesfullywith thegivenservicetask.2

Thehigh level plannerneedsa mapof the environmentto scheduleanddistributethetasks.Althoughtheenvironmentis known a-priori, its structurecanalsointroducesomehazardsin additionto the onescommingfrom the sensorsdueto similaritiesinstructure,steadychangesin theenvironmentandmachineunfriendlypaths,e.g.windingcorridors,toomany doorsin asmallarea,etc.Theseaspectsaresupportedby themapoftheenvironmentthatunderlaystheplanner. Thedesignednavigationschemehasto berobustenoughto recoverfrom erroneouspathplanningonits own withoutreschedulingthe completepath e.g. by manouevering autonomouslyout of a dead-end.Last, butnot leastsomeof themissinginformationgiventhroughincompletecommandsto the

2 A great number of different planners and testbedscan be found under http://eksl-www.cs.umass.edu/planning-resources.html.

systemcanbe compensatedthroughapriori knowledgeaboutthe environmentby thesystemitself.

The presentedexperimentis the first stepin a larger projectdealingwith servicerobots.In this experimentwe have mainly concentratedon thefirst threefeatures:au-tonomy, environment,andhumaninterface.Robotcooperationhasbeentakeninto con-siderationwhile designingtheplannerandthemapof theenvironment,thuspreparingthegroundwork for theextensionto agroupof servicerobots.

The chapteris structuredas follows. Sectiontwo presentsthe architectureof theimplementedsystem.We startby giving thereasonfor a fuzzy approach,by introduc-ing theproblemto besolvedandby giving theapplicationconstraints.Next a generaloverview of the architectureis introduced.In the following two sectionswe describedetailsof the architecture.In sectionthreewe describethe plannerandthe useof theenvironmentmap.In sectionfour we presentthenavigationblock andtherelatedlocalperceptionof the environment.Sectionfive presentsthe experimentalresultsconcen-trating on the monitoringandcommunicationcapabilitesof the system.We concludeourpresentationwith somefinal remarksandanoutlookon futureactivities.

2 The Ar chitecture of MORIA

2.1 Why A Fuzzy Approach?

As mentionedin theintroduction,servicerobotsoperatingin structuredreal-world en-vironmentsrequiretheability to copewith uncertain,incompleteandapproximateen-vironmentalinformationin realtime. In addition,astherobotshave to executehumanscommandsissuedby personnelwith differentlevelsof qualification,uncertaintiesdueto incompleteor missinginformationhave to betackled.Theproposedrecurrentfuzzysystem(RFS)that implementsanautonomousintelligentsystemtakesinto considera-tion all theseuncertainites.Beforediscussingtheproposedapproachwe would like todescribein moredetail theuncertaintiesexisiting in theenvironmentandtheir impacton therobot.

Thehighuncertaintyexistingin theenvironmentis themainchallengein thecontrolandbehaviour of autonomousnavigation.Thetypesof uncertaintycanbeclassifiedintosensorydependentandinformationdependent.Sensorydependentuncertaintycausedby the low reliability of thesensorscanbereducedthroughsensoryfusionandby us-ing additional,a-priori knowledgeexisting within the system.Informationdependentuncertaintyoccurswhenthemapof theenvironment- a-priori known structuredenvi-ronment- doesn’t fit reality dueto somedynamicchanges.Themapcanbeconsideredincomplete.Themissinginformationis dueto:

– anincompleteplanof theenvironment(missingdynamicobjects).– steadychangesin theenvironment(appearanceanddisappearanceof dynamicob-

jects);e.g.anobjectblockscertainpathfor agivenperiodof time.– transientperturbationin theenvironment(moving object- humansor robots)that

distortenvironmentalmeasuresby occludingfixedobjectslike walls.

Thesevariousuncertaintiescancausethefollowing typesof errors:

– wrongdescriptionof the local environment;theerrorcanpropagateinto thenavi-gationstrategy andstartawrongmanoeuvrer

– wronglocalisationof therobotin its own internalmap;suchanerrorcouldtriggeranotherpathestimationcausingtherobotto turnandstartlookingfor its destinationusinganother, generallylongerpath.

Thefirst typeof errorhasalocaleffect.Thelocalnavigationof therobotis changedbutthepathfollowedto reachthegiventargetremainsthesame.Thesecondtypeof errorhasa largerimpacton thesystem.It introducesa globalerror jeopardisingthesuccessof themission(task).

If we summarizetheuncertaintiesdescribedabove it canbestatedthatknowledgeof the environmentis inherentlypartial andapproximate.This hasan impacton boththe naviagtion andthe taskplanningof the autonomoussystem.Sensingis noisy, thedynamicsof the environmentcanonly be partially predictedand the hard-andsoft-waretasksexecutedby the system(robot) arenot completelyreliable[16]. Classicalpathplanningapproacheshave beencriticised for not beingable to copeadequatelywith this situation3. Traditionalapproachesto mobilerobotnavigationhave dealtwiththeserequirementsby usingcomputationallyintensiveplanningalgorithmsandexplicitpre-determinedworld models[18]. This is not necessarilya drawbackfor fixed-basemanipulatorsbut it is a problemfor mobile robotsfor which computationalresourcesmustbecarriedonboard.

In ouropinionthefirst partof theansweris not to compensatefor theuncertaintyata givenprocessinglevel throughhigherprecisionin the following processstep,but tomake decisionsbasedon qualitative measurementsthatgive a certainrangeof valueswith agoodprecision.By usingqualitativedecisionmakingtheprecisionof theneededinformationdecreasesandthusbettermatchesthe input informationrequestedby thecontrol system.Thisqualitative decisionmakingcanbesupportedthrougha fuzzy ap-proachwherethedecisionfor thebehaviour is notdecidedthroughacrispstructurebutthroughanevaluationof thebestpossibilitywithin a givenrange.As will beshown insectionfour ournavigationblockutilisesthisdecisionmakingapproachby introducingfuzzystatevariables(FSV).

Thesecondpartof theansweris greatautonomyof thesystemrelatedto its abilityto planapathbetweentwo givenpointsunderthegivenconditions.A classicalplanningstrategy needsanaccuratemapof theenvironment.While qualitativeenvironmentalin-formationis enoughfor navigation(collision avoidance),this typeof informationis ofcoursenot enoughfor a precisetopologicalmapof the environment.Let’s analyseinmoredetail if a precisegeometricmapof theenvironmentis needed.Many pathplan-ning approachesdescribethepathasa list of coordinates.As long astheenvironmentis restrictedto aroom,theamountof memoryneededis manageable.But if wethink oftheservicerobotsdriving througha building thememoryneededfor suchanaccuratemapincreasesvery quickly andtheprocessingtime linearly with it. As alreadyshownby Saffiotti [11] a fuzzy basedcontrollerhastheadvantagethat the intuitive natureofcollision-freenavigationcanbeeasilymodelledusinglinguistic terminology(e.g.nextright insteadof 89� 25o onpreciseradius)[14]. Thisimpliesthatthelow-levelbehaviour3 Saffiotti [17] givesaninterestingglanceat the“planningvs. reactivity debate”

information

explorationcommands for

module

cartographicgenerationmap

path planning

planner module

map planner

location

path

navigator

sensing controlactions

command list

topologic map

navigator

(task generator)

path planner

local env. inf.

Fig.1. SystemArchitectureof MORIA

approachallows thestoringof informationat a higherlevel of abstraction.If this is so,a fusionbetweenplanningandnavigationcanbedoneby combininga high level goalwith anautonomousnavigationwithin this goal.This combinationbetweenhigh levelgoalandlow level topologicmap-basednavigationcanalsobesupportedthroughthefuzzy statevariablesintroduced.Thusthe planneris alsobasedon a topologicalmapthatdoesn’t keepthehigh accuracy of otherconventionalplanners,but thegoalof themissionis still reached.Usingthisapproachwehaveextendedthefuzzyconceptsto theplanninglevel too. Therobotwill alwaysreachits globalgoalbut nobodycanpredicttheexactcoordinatesof its pathexceptwhenthey aregivenasanintermediategoal.

By usinga fuzzy approachboth for planningand for navigation our autonomoussystembehavesat the sametime in a reactive anda goal orientedmanner. The twodifferentbehavioursareblendedtogetherdependenton theenvironmentalinformationandtheinternalfuzzy statevariablesTheobtainedsystembehaviour althoughbuilt upusingonly a few blocksis alwayschangingandadaptingin anadequatemannerto theenvironment.Theproposedblockarchitectureof thesystemis presentedin Fig 1.

2.2 Problemspecification

MORIA (Fig 2) is designedasa transportsystemfor warehousesandoffices.Thema-jor hardconstraintin our designspecificationis thecostof thecontrolarchitecture.Tobecompetitivein themarket,our industrialpartnerrequestedasolutionfor autonomousnavigation thatwould costlessthanUSD 6.000.This of coursehadan impacton thechosensensors.Themaintaskof thesystemis to drivewithin anbuilding from A to B

F

MORIA

B

BR

RL

FRFL

BL

MORIA

1 softbumper

2 wheel encoders

(2 motors: direction and angle) 1 driving front wheel

2 rear wheels

touch bumper

emergency button

1 camera (optional)

horn

infrared-link

optical alarm

8 sonor sensors

PC 486/33

battery status

status monitor

a) b)

Fig.2. Theequipmentof Moria

with thehelpof a map.Thesystemshouldfind thebestpathdependingon thetypeofload.In caseof anunforseenobstacleor objecttemporarilyblockingthepath,thesys-temshouldbeableto recalculatethepathwithout needingthesupportof a supervisor.

Navigation is basedon 8 sonarsensorsandtwo wheel-encoders.The distributionof the sonarsensorson the systemis given in Fig. 2b. The alarmsystemis triggeredby touchsensorsintegratedin thefrontal, lateralandrearledgebumpers.Thewarningsystemconsistsof ahornandthreeflashinglights.Thelightsareusedto give thestatusof thevehicleusingcodedsignals:X-X-X for forward,X-X-0 for left turn, 0-X-X forright turn,etc.Thehigh-levelplannercanbelocatedeitherontherobotor onanexternalcomputer. In the secondcasean infraredlink is usedfor the communicationbetweenplannerand navigator. This communicationchannelis important if a pool of robotsare working in the sameenvironments.Eachrobot hasa navigator and one plannerdistributesthe taskandcalculatesthe paths.The infraredlink hasa rangeof 40m,sofor abuilding severalaccesslinks areneeded.

The interfacefor the useris a keyboard,a monitor anda joystick. The joystick isusedfor manualdriving of therobot.Thekeyboardcanbelocatedeitherontherobotoron anothercomputer. Thesameinfraredlink is usedfor remotecontrol.Themonitor, a2D mapof theenvironment,is onaremotesystemandreceivestheactualdatafrom therobot via the infraredlink. This approachenablesus to useoneinterfacebetweenthelow-level navigatorandthehigh-level commanders(planneror human).Thenavigatordoesn‘tseethe differencebetweenthe two modes.The commandsgiven by humansareof type “next left”, etc.The systemcanbe enhancedwith an optionalcameraforadditionaltasks,e.g.visualisationin a 3D virtual reality (VR).

2.3 Proposedapproach

Theabovementionedspecificationrequiresanautonomousreaction(action)in thefol-lowing situations:

– giventhemapof theenvironmentthesystemis ableto find a pathbetweena givensourceA and a given destinationB underspecialconstraints.For example“thenumberof turnsshouldbe aslow aspossible”or “the corridor shouldbe alwayslargerthanagivendistance”,etc.

– thesystemshouldavoid collisionwhile driving to its destinationwithout losingthepredefinedpath

– thesystemshouldidentify dead-ends,find thebestway out andif necessaryrecal-culatethebestpathfrom thegivenposition.

If weanalysethesebehaviourpatternsandtakeintoconsiderationtheirimpactrangewecandivide theminto globalandlocal.Actionswith globalimpactarerelatedto pathfinding, localisationof therobot in themapandupdatingthemapof theenvironment.The plannersupportsall thesetasks.For thesetasksa soft real time processingcon-straintis necessary. Theplannerperformingthesetasksis far-sightedandgoal-oriented.

The local actionsof the systemarerelatedto collision avoidanceandto adapttounforseensituationsappearingin the nearneighbourhood.Thesenavigation tasksre-quirehardreal-timeprocessing.Thenavigatorperformingthesetasksis short-sightedandreactive.Unexpectedsituationsaremainlycausedbymoving objects.Theseobjects(events)arelocal andcanappearanddissappearwithin anenvironment.Their impactcouldcreatea mismatchbetweentheinternalmapof theenvironmentandreality, thushavealsoaglobalimpact.Thesystemwill first reactvia thenavigatorandadaptlocallyin real-timeto thesituation.Theextractedinformationwill bepassedto theplanner. Ifa global impactis detectedthis changewill beprocessedby theplanner. By choosinga reactivenavigatoranda goal-orientedplannerthesystemis ableto copewith knownandunknown situationwhile keepingtheimposedtime constraints.

Theblock architectureof thesystemis presentedin Fig. 1. If a taskis givento thesystemthentheplannerstartsthefollowing process:

– the startandendpoint aredetectedin the internalmap;additionalinformationisrequestedfrom theinternalknowledgebase

– thepathplannerestimatesthebestpathfrom A to B– the taskgeneratortransformsthe path into a list of linguistic commandsof type

“straightahead”,“next left”, etc.

The navigator hasonly “one commandin mind” at a time and tries to fulfil it whileavoiding obstacleson its way. This meansthat thefollowing processingstepsareiniti-atedby thenavigator:

– thesensorialinformationis fusedandevaluated– the executedcommandis acknowledgedso the plannercantrigger the next com-

mand– atopologicaldescriptionof thelocalenvironmentis generatedandsentto theplan-

nerto up-datethemap

At thefirst glancethesystembehavesasexpected.Theplannertakescareof globalissuesandhasa longerestimationtime for this while thenavigatorreactsquickly onlyto sensorialinformationandkeepstheimposedreal-timeconstraint.

branchreactivity

statedetection

turningreactivity

speedadjustment

manualcontrol

accidentreactivity

blockedcorridorreactivity

reactivitybasic

state 3: moving course;state 2: right junctionstate 1: left junction

state 4: moving direction;state 5: accident danger;state 6: narrow passage

left, straight, rightbackward or forwardleft or right

state 7: blocked corridor

Sens.

Cmd

FSV

Mot. Ctrl.

FSV

a) b)

Fig.3. Detail of therecurrentfuzzy system;a) fuzzy behaviour blocks;b) fuzzy statevariables.Left junction,Rightjunctionandmoving directionhavecrisp,non-overlappingmembershipfunc-tionsandtheotherstatesGaussianlike,overlappingMFs.

But thisapproachis notsufficient for solvingconflictingproblemsbetweennaviga-tor andplanneror reactingadequatelyto unexpectedsituations.A simpleexampleillus-tratespossibleconflicts.Supposesomebodyputsa tablein thecorridor (largeenoughto block it for therobot).This informationis not in themap,thereforethepathplannergivesacommandto gostraightthroughthecorridorignoringthechange.Thelocalper-ceptualblockof thenavigatoridentifiestheblockedcorridor. For awhile thecommandcomingfrom theplannerhasto bepost-ponedandthebehaviour “basicreactivity” hasto bechangedinto “blockedcorridorreactivity” of thenavigator(seeFig.3a).Theplan-nerhasto beinformedabouttheoccurrenceof thespecialsituation.Fromthepoint ofview of taskdistribution, the navigator is now the masteruntil the unexpectedsitua-tion is solved.To handlesuchsituationswe introducedfuzzy statevariables(FSV).Atall times thesevariablesidentify the local situationbasedon the sensorialinput, thecurrentglobalcommandandtheprevious“state-of-mind”of therobot(previousfuzzystate)(seeFig.3b).If weblendtheseinputstogether(command,sensorialinput,andthefuzzystatevariables)into onefuzzy inferenceengine,we havea goalorientedreactiveautonomoussystem(seeFig 7). The modulardesignapproachhasleadto a recurrentfuzzysystemimplemention[19]of theautonomousbehaviour.

2.4 A Recurrent Fuzzy System

A (first order)recurrentfuzzy system(RFS)is characterisedby rulesin which oneormorevariablesappearboth in thepremiseandconsequentparts,like x in: IF x(t-1) isAkjANDu

�t � 1� is Bkj THEN x(t) is Ckj , whereAkj representsthemembershipfunctions

codingthe linguistic termk associatedto thevariablej [19]. A recurrentfuzzy systemmeansa fuzzy systemwith inferential changing,an approachvery typical of fuzzyreasoningor fuzzyexpertsystems,butveryinfrequentlyin fuzzycontrolwhereoneshotinput/outputstructuresarecommonplace.An n-th-orderRFSmayalsocontaininternalvariablesfrom several (n) differentblocksof rules.The fuzzy internalvariablesalsohave their own temporaldynamics.Fig. 5 shows two examplesof fuzzy statevariablesfor Moria.

turning right turning left left accidental right accidentalright junctionright junction left junction

blocked corridorblocked corridorrotation right rotation left

blocked corridorturning right

blocked corridorturning left

blocked corridorright junction detected

advanced turningblocked corridorleft junction detected

a) b) c) d) e) f) g)

h) i) j) k) l) m)

n) o) p) r)

s) t) u)

left near right near ileft near right near in front near right transitleft transit narrowcorridor blocked corridor

Fig.4. LocalPerceptualStateandthenamesof therelatedfuzzy statevariables

In contrastto RFS,most autonomousrobotshave standardFS without memory.They act or reactonly basedon the actualsensorialinput. Ambiguousstatessuchas“narrow passage”or ”accidentdanger”(Fig. 4 f,p-r) aredifficult to recogniseimmedi-ately whentherobot is in locomotion.By introducingfuzzy statevariables(FSV) weaddeda local memoryto thenavigationprocessandgave the systemthecapabilitytoreactin a temporalydynamicway to complicatedsituations(Fig. 4n-u) basedon thenearpast.Another advantageis the smoothpassagefrom one behaviour to the nextthroughthe FSV. The different strategies are smoothlyblendedtogether(in a fuzzymanner)dependingon thedegreeof thebelongingof theenvironmentalconditionstoa definedstate.Not all statesarefuzzy, somearecrisplike “left junction”. An exampleof theusedfuzzy controlleralgorithmis givenin Fig. 6. Thecalculationof thecenterof gravity for theresultfunctionspeedis themajorbottleneckof thefuzzy algorithm.Therefore,a modifiedresultfunctioncalculation[20] is suggested,in which thecenterof areaMi andtheareaAi of a membershipfunctionarecalculatedbeforerun time (m

0

0.2

0.4

0.6

0.8

1

-2 0 2 4 6 8 10

BLCORRI

rs

lbrrbrlbrrrbrrrot

laccracc

0

0.2

0.4

0.6

0.8

1

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

moving direction

forwardbackward

Fig.5. Exampleof overlappingandnon-overlappingmembershipfunctionsfor the fuzzy statevariableblockedcorridorandmoving direction

if F is large and then

then

and thenif F is medium FL is medium

0m 2m 4m

large1

0

0m 2m 4m

medium1

00m 2m 4m

medium1

0

and

small

0m 2m 4m

1

0

if F is small COG

0

0

1 m/s

1 m/s

1 m/s

1

1

1

0m 2m 4m

medium1

0

FL is medium

0m 2m 4m

medium1

0

FL is medium SPEED is fast

SPEED is slow

SPEED is medium

0

Combined inferenceresults.

1

0 1 m/s

fast

medium

slow

Fig.6. Exampleof thefuzzy controlleralgorithm.F: front sensor. FL: front left

= numberof rules):

Ai ��� speed� e� de� Mi ��� e speed� e� deCOGI �

m∑

i 1ωi � Mim∑

i 1ωi � Ai(1)

The setof fuzzy statevariablesrepresentsa local perceptualandactionmemoryof thefuzzy controller(FC) (Fig. 3). Sincestandardfuzzy membershipfunctionsmayimplementedbothascrispor fuzzyvalues,statescouldalsobecrispor fuzzy.

Furthermorea fuzzy statevariablerepresentsa landmarkfor the topologicalmap.Let’s look at a moredetailedrepresentationof thesystemarchitecture(Fig. 1).

Theplannergeneratesa list of commands(seesection3). Only onecommandat atime is sentto theFRBS.Thenext oneis triggeredby thesensorialinformationcoming

from the Local PerceptualSpaceandthe FSV. For exampleif the commandis “nextleft” andtheFSV =1 (Fig. 3b) thiswill triggerthenext commandfrom theplannerlist.

The navigator hasseveral behaviour modules(Fig. 3a) eachbeing a fuzzy rule-base.If the situation i occurs,the correspondingfuzzy rulesareselectedthroughanadditionalfuzzy rule (context rule), andselectstheconditionvariablei. In contrasttothe approachpresentedby [13] our vehicledoesn’t needthe definition of specialruleweightsto changefrom onenavigationstateto thenext. This is doneby thefiring valueof thecontext rulesandwhicharenotseta-priori.Furthermorethenew fuzzystatevari-ablesat thenext timestept+1 is calculatedfrom thesensorialinputandtheactualfuzzystatevariableat time stept. As anexample,if thepremise“IF left sensor= very smallAND front sensor= default AND right sensor= very small” is activatedthena narrowcorridoris detectedandtherobothasto driveslowly. UndertheseconditionsFSV5 and6 areactivated(Fig 3b).Thefuzzy ruleshavethefollowing form for themotorcontrol:

IF COMMAND is CiAND SENSOR(1)is Ih1 ... AND SENSOR(m)is IhmTHEN SPEEDis Sj AND ANGLE is A jorIF COMMAND is CiAND STATE(1) is ST1l1 ... AND STATE(n) is STnlnAND SENSOR(1)is Ih1 ... AND SENSOR(m)is IhmTHEN SPEEDis Sj AND ANGLE is A j

For theestimationof theactualFSV theform is:

IF COMMAND is CiAND STATE(1) is ST1l1 ... AND STATE(n) is STnlnAND SENSOR(1)is Ih1 ... AND SENSOR(m)is IhmTHEN STATE(1) is ST1l1 ... AND STATE(n) is STnln

Thecollision avoidancestrategy basedon theserulesis explainedin moredetail insection4. Theimplementedrecurrentfuzzy systemhas29 inputs:7 FSV’s,8 sensorialinputs,1 commandinput,8 sensorialderivativesand5 FSVderivatives.Thenumberofoutputsis 14: 2 for motor control and12 FSV’s. The implementedrule-basehas352rules.Theserulescanbe divided into blockscorrespondingto the variousbehaviourspresentedin Fig. 3. Theadvantagesof theapproachcanbesummedup:

– keepthemodulardesignof high-level plannerandlow-level navigator

– haveasmoothpassagefrom goal-orientedto reactivebehaviour withoutconflicts

– keepa modularstructureof the autonomousbehaviour by implementingseveralbehaviour blocksseparately

– extendto new behaviour typeseasilyby addingnew fuzzy statevariablesandnewbehaviour blocks.

sonar sensor 1 state 1

state 2Hysteresis!speed

angle

fuzzy

fuzzy

controller

statemachine

state variable n

state variable 1 state variable 1

state variable n

sonar sensor m

sonar sensor 1command

feedback

local perceptual memory

sonar sensor m

Fig.7. A RecurrentFuzzySystem.With thehelpof the internalfuzzy statevariableshysteresiscouldbe realized.Theconceptof fuzzy statevariablemay containboth:crispandfuzzy mem-bershipfunctions

3 Goal-orientedPlanner

3.1 Path Planner

Mostpathplanningapproachesarebasedon highprecision,metricapproachesandde-mandveryreliableandaccuratesensordevicesaswell asgreatcomputationalpower. Inaddition,thesemethodshavedifficultiesduringnavigationin complex anddynamicallychangingenvironmentswhereunknown obstaclesappearon an a-priori plannedpath[17,14]. Theconventionalpathplanningapproachesfor mobilerobotsmaybedividedinto two categories.Oneis globalpathplanningbasedona-prioricompleteinformationabouttheenvironment.Theotheris thelocalpathplanningbasedonsensorinformationin uncertainenvironmentwheresize,shapeandlocationof obstaclesareunknown.

In ourapproachthenavigatoris analogousto adriverandaplannerto aco-driverofthesystem.Theplanneris in chargeof themapandof findingnew routes(paths)whenunexpectedobstacleappear. It utilisesa topologicalmapof the network of corridorsandhasthe path-findingsophisticationto plana routefor the robot.Theplannerusesonly computedstatevariablesof the navigator combinedwith fuzzy metric informa-tion. The fuzzy statevariablescontaininformationaboutactionsof the robot, relativefuzzy distancessincethe last fuzzy stateinteractionandinformationaccordingto thecurrentsituation.Thetaskgenerator(seeFig.1) transformsthefoundpathinto a list ofcommands(Table1). Fig. 8 showsanexample.

In caseswhereobstaclessuddenlyappear, fastreactionof therobotis needed.SinceMORIA driveswith amaximumspeedof 1 � 0ms therobotmayhaveto makeahardbreakand/orturn.To avoid a jerky driving stylewe implementedsomeadditionalcommands(4-7 in table1) wherethe navigator executesimmediatelywithout exploring the best

0

1000

2000

3000

4000

0 1000 2000 3000 4000

[cm

]�

[cm]

a

f

d

e

bc

Fig.8. Path-planningandthe correspondingcommandlist: next left (a); next right (b); straightahead(c); next left (d); straightahead(e); stop(f).

next environmentalpossibility.

Obviously, metric informationaboutprecisedistancesarenot necessary. Themazeof corridorscan be representedas a graph.The path planningalgorithm acts like avectoredgraphsearch.To find thebestrouteis anNP completeproblem.We restrictedthe searchby reducingthe maximumnumberof corridorsto onehundred.This is arealisticfigureevenif thebuilding is very complex with a largenumberof floors.Thesearchcan be donerecurrentlyfor eachfloor. A detailedpresentationof the searchalgorithmcanbefoundin [15].

3.2 Map of the envir onment

Geometricapproachesusedistanceinformationfrom severalsensorsto build geometricmapsmostly basedon probability theory e.g. [21]. Theseapproachesneeda lot ofpreciseinformationandhave difficulties in distinguishingbetweenstaticanddynamicobjects.An off-line generationof a topologicalmapout of the geometricinformationcanleadto errors.Explorationasagraphconstructionwithoutdistanceinformation,e.g.[22], requiresfrom therobot to dropdistinctmarkers.Our approachexploresdifferentlevels of maprepresentationsimultaneouslywithout droppingdistinct markers.Eachmaptypehaslimitationsandis thereforeusedfor partof theplannertasks.

Theservicerobotsactin astructureda-prioriknown environment.Thestaticmapoftheenvironmentis usuallyavailablein acomputerreadableformat.In ourcaseit is readautomaticallyinto thesystem.Part of it is shown in Fig. 9. This mapis thenconvertedinto atopologicalgraphthatis usedto searchpaths.Onthegraphlevel,wetry to extract

Table 1. List of commandsunderstoodby thenavigator. They areimplementedasGaussianlikemembershipfunctionswith y x��� 0 in thegiveninterval andy x��� 1 in themiddleof theinterval

Commands FuzzynumberStop(s) 0 ��� � 0 � 5 � 0 � 5�

StraightAhead(a) 1 ��� 0 � 5 � 1 � 5�Take Next Left (l) 2 ��� 1 � 5 � 2 � 5�

Take Next Right (r) 3 ��� 2 � 5 � 3 � 5�GoBackward(b) 4 ��� 3 � 5 � 4 � 5�

TurnLeft Immediately(li) 5 ��� 4 � 5 � 5 � 5�TurnRight Immediately(ri) 6 ��� 5 � 5 � 6 � 5�

GoForward(f) 7 ��� 6 � 5 � 7 � 5�ChangeDirection(cd) 8 ��� 7 � 5 � 8 � 5�

new: LastLeft (ll) 9 ��� 8 � 5 � 9 � 5�new: LastRight (lr) 10 ��� 9 � 5 � 10� 5�

����� ���� ����! ����" �#�%$ �����&�'

����( �#�%) �#�%* ���%+ �#�%(&-, ./0123 4156.79817.

'�:;:;= ?;:;

0

920

1840

2760

0 920 1840

[cm

]B

[cm]

Fig.10.Map generatedby thesensorialinformation

4 Reactive Navigator

The recurrentfuzzy inferenceengineof the systemoperatesbasedon the fuzzy inputvariables:command,sensorialinformation, the FSV, the speedanddirection (angle)of the motor, the angleof the front wheel,and the new fuzzy states.The avoidanceprocedureis basedon a reactive control loop betweensensorsandmotorsby slightlychangingthedriving angleandreducingthemotorspeed.

The“basicreactive behaviour” of thefuzzy controllerusesonly threesensors.Fordriving forward the sensorsFL, F, FR asshown in Fig. 12. For driving backward thecorrespondingsymmetriconesBL, B, BR, areused.Thelateralsonarsensorsleft L andright R areusedfor dockingor driving parallelto awall by evaluatingBL, L, FL for theleft wall andBR, R FR for theright one.Therangeof valuesof eachsensoris dividedinto threelinguistic terms:“verynear”,“near”, “standard”.For eachcombinationof thethreemembershipfunctionsa fuzzy rule is given (seeTable2). The rule baseof thebasicbehaviour is completelydefined(27 rules).

The systemoutput valuesfor speedand rotation angle(direction) are estimatedthroughsuperpositionof thevariousforcesactingin thedecisionmechanism:sensorialforces,commandforces,FSV forces.Let’s considerfirst the forcesintroducedby thesensorialinputasshown in Fig. 12.Thefiring rule will be:

IF FRONT-SENSORis very-nearAND FRONT-LEFT-SENSORis nearAND FRONT-RIGHT-SENSORis very-nearTHEN SPEEDis (positive)-small,ANGLE is negative-small

Thevehiclewill startturningto theleft astheFL sensoris only nearcomparedwiththeothertwo values.If we addthecommandcomingfrom theplannerto theinferenceprocessthebehaviour couldbechanged:

case1:

0

920

1840

2760

0 920 1840

[cm

]B

[cm]

Fig.11.Topologicalmapbuilt on sensorialinformation

Table 2. Basicreactivity rule base

left center right angle speed

normal normal normal zero highnormal normal small s. left med.normal normal v. small left smallnormal small normal s. left med.normal v. small normal left smallsmall normal normal s. right med.v. small normal normal right small. . . . .. . . . .v. small v. small v. small zero zero

IF COMMAND is straight-aheadTHEN SPEEDis default,ANGLE is zero.

case2:IF COMMAND is left-rotationTHEN SPEEDis very-small,ANGLE is negative-very-large.

If case1 is active thecommandforcesdon’t changethemotorcontrolrequestedby thesensors.If case2 is active thefinal outputwill be:

...THEN SPEEDis almost-very-small,ANGLE is negative-medium

If we integrate(superimpose)thecommandrulesandthesensorialrulesthenthe rule

MORIA

FLob

ject t

o th

e lef

t (wa

ll)

FL

F

FR

very

nea

r

near

very neal

near

very near near

object to the right (wall)

stand

ard

standardl

standardFR

Fobject in front (wall)

Fig.12.Fuzzydirectionforcesfor a blockedcorridorandtheinputmembershipfunctions

for case1 is:

IF COMMAND is straight-aheadAND IF FRONT-SENSORis very-nearAND FRONT-LEFT-SENSORis nearANDFRONT-RIGHT-SENSORis very-nearTHEN SPEEDis positive-small,ANGLE is negative-small

Now let’s introducetheforceof theFSV. Their impacton thebehaviour canbebestexplainedwhena blockedcorridorappears.Thebasicreactive rule basewill decreasethe speedandslowly increasethe turningangleto the left (negative large).The fuzzystatedetectionwill fire thefollowing rule:

IF FRONT-SENSORis very-nearAND FRONT-LEFT-SENSORis very-nearANDFRONT-RIGHT-SENSORis very-nearTHEN FSVenvC is blocked-corridor.

This fuzzy statewill changethe behaviour of the systemfrom basic to “block-reactivity”. Thefiring rule is:

Fig.13.3D graphicaluserinterfacewhich shows a planview of therobot’s progressthroughthetestenvironment.

IF FRONT-SENSORis very-nearAND FRONT-LEFT-SENSORis nearAND FRONT-RIGHT-SENSORis very-nearAND FSVenvC is blocked-corridorTHEN FSVdirection is backwards.

Thesystemstartsthe recoveringmanoeveur from a blockedcorridorandthe FSVinformstheplannerabouttheidentifiedstate.Theplannercaneitherkeepthelastcom-mandif the robot wason explorationtour or recalculatea new path.As presentedinsection2.3 the behavioural blocksarefired in parallelandall inputsareestimatedatoncethusproducingthedescribedrecurrentfuzzysystem.

5 Implementation

5.1 Designand Simulation Envir onment

This projectusedthe designandsimulationenvironmentthat wasinitially developedfor the digitally implementedfuzzy control systemreportedin [23]. The modellingsystemFunnyLab [24] wasusedto createandedit the membershipfunctionsandtherule base.On a PC 486 / 33 MHz the generatedfuzzy engineprocessesat a speedofmorethan2 MFIPS(million fuzzyinferencepersecond)includingdefuzzification[25].Theknowledgebasefile producedbyFunnyLabwasreadinto asimulationenvironmentdevelopedin-houseat GMD, which includesnot only a fuzzy inferenceengine,butalsoa motionsimulatorwhich incorporatesa modelof thereal,measureddynamicsofthe robot MORIA. A graphicaluserinterfacewhich shows a plan view of the robot’sprogressthroughthetestenvironment,displaysthevaluesof certainkey parameters.Itallows the instantaneousentryof the plannercommandsgiven in table1. The motionsimulatorhasbeendemonstratedto produceanaccuraterepresentationof thedynamicsof MORIA during earlierdevelopmentwork, andallows a realisticassessmentof theeffectsof thefuzzy controller, giventhat it includestheeffectsof theinertiaandfiniteresponsetime of therobot.

Theuserinterfaceincludesan additionalfeaturewhich allows single-stepping.Ateachstepit is possibleto watchtheactivatedfuzzy rules,thesensorialoutput,andthe

valuesof thefuzzyvariables.Thisfeatureis apowerful tool for examiningtheoperationof the rule baseover critical stretchesof travel. This facilitatestherapid identificationof theactualfiring rule, thetuningof thebehaviour blocks,or thefuzzystatevariables.

5.2 The Robot MORIA

The vehicleMORIA is a mobile device with dimensionsof 157cmx 90cm x 73cm(L,W,H). The weight of the vehicle is 400kgand it hasa payloadcapacityof 150kg(Fig. 14). The vehicle is driven by two motorsacting on a single wheel situatedinthefront of thevehicle,onemotorthatis reversibleprovidesthedriving torque,andtheotheronesteersthevehicleby changingtheorientationof thedriving wheel.Dueto thisconstruction,the forwardandbackwarddriving strategiesaredifferent.It is equippedwith 8 ultrasonicsensorsthathavea rangebetween50 - 400cm.

Fig.14.MORIA at theHannover Fair’96.

Computationalcapabilitiesof MORIA consistof anindustrialPC(486/66MHz, 16MBytes)with extendedI/O possibilities.ThePCboardcollectstheoutputof thesonarsensorsandcontrolsthetwo motors.A communicationlink to othermobileplatformsor remotecomputersis availablevia anon-boardinfraredsensor.

5.3 Experimental results

The vehiclewas testedin the corridorsof our researchinstitution andseveral publicpresentationsweremadeduring theHannover Fair’96 andGMD OpenHousedaysinthe last 3 yearse.g.during the Hannover Fair the robot drivesoneweekwithout anaccidentandonly two deadlocks.The first stageof the implementation,the low levelreactivity andhigh-level plannerwerepresentedat the IEEE FUZZ/IFES’95whereittook the“IntelligenceAward”.

Theexperimentalresultsprovednot only therobustness(∑ of userhelps/ (numberof operatinghours)) of the approachbut alsoshowed the simpleprogrammabilityofthe systemthroughthe separatelydesignedbehaviour blocks.Dependingon the testenvironment(width of thecorridor) thesamerule basewith differentdefinitionof thelinguistic termsof thesensorialinformationwassuccessfully. Experimentswith differ-entbehaviour typeswerealsotestedandaneasydownloadmechanismsupportedthisadaptivity. Thecomplexity of theenvironmentwasincreasedevery year. At theendofour experimentswith MORIA thesystemwasableto drive throughthevariousfloorsusingthebehaviour blocksandthenumberof rulespresentedin this paper.

The major gain from theseexperimentswasthe experiencewe acquiredin speci-fying the featuresof a servicerobot for buildings.Basedon this experiencewe havedefineda new typeof robot(a groupof three)which wearecurrentlystuddingThebe-haviour developedfor MORIA wasportedandthenew mechanicalconstructionof thesystemwasdesignedto betterfit our localenvironment.Thenew family of robotshavea greaterrangeof operation,e.g.they canusetheelevator. Themonitoringcapabilitieswereenhancedto permitgoal-orientednavigation for all threerobotsto beperformedon aremotecomputer.

6 Conclusionand futur e work

This paperdescribeda low-level navigator anda high-level plannerthat were imple-mentedasa fuzzy recurrentsystem.By introducingthe fuzzy statevariableswe wereableto drive andcontrol therobot througha complex environmentandthesystemre-actedautonomouslyto unexpectedsituationssuchasmoving or dynamicobjects.Ourexperimenthasalsoshown thatautonomoussystemscanbecontrolledwith humanlikebehavioural commandsthussupportinga humanfriendly interface.A new generationof robotsis currentlybeingimplementedin theARIADNE projectbasedon this expe-rience.Themaingoalof our researchwill be thecooperationof therobotsundertest.This featurewill be basedon an extensionof the global plannerhousedin a remotesystemandthereactivenavigatorpresenton theautonomousplatform.

AcknowledgementJ. Huser, ShuweiGuo,StefanVirsik, MarkusEderandJ. Wehkingcontributed to the developmentof MORIA by implementingseveral modules.TZNUnterlüß(Germany) supportedpartof thiswork. JulianKolodko andthereviewermakesuggestionfor improving this paper. Our thanksto all of them.

References

1. R. D. Schraftand H. Volz, Serviceroboter, Innovative Technik in Dienstleistungund Ver-sorgung. Berlin, Heidelberg, New York: SpringerVerlag,1996.

2. E. Prassler, R. Dillmann, andH.-B. K. (Hrsg.),Robotikin Deutschland: Lehre, ForschungundEntwicklung. Aachen:Shaker Verlag,1998.

3. S.Thrun,“Learningmetric-topologicalmapsfor indoormobilerobotnavigation,” ArtificialIntelligence, vol. vol. 99,pp.21–71,1998.

4. R. BajcsyandJ. Kosecka,“The problemof signalandsymbol integration:A studyof co-operative mobile autonomousagentsbehaviors,” in Proceedingsfrom DAGM SymposiumBielefeld,Germany, September, 1995.

5. K. SekiyamaandT. Fukuda,“Modeling andcontrolling of groupbehavior basedon self-organizingprinciple,” in Proc.of the1996IEEE InternationalConferenceon RoboticsandAutomation(ICRA’96), vol. 2, pp.1407–1412,1996.

6. A. Birk, “Behavior-basedrobotics,its scopeandits prospects,” in The24thAnnualConfer-enceof theIEEE IndustrialElectronicsSociety, IECON’98, 1998.

7. H. Surmann,A. Kanstein,andK. Goser, “Self-OrganizingandGeneticAlgorithms for anAutomatic Designof Fuzzy Control and DecisionSystems,” in Proceedingsof the FirstEuropeanCongresson Fuzzyand IntelligentTechnologies,EUFIT’93, Aachen, pp. 1097–1104,7. - 10.09.1993.

8. C.C.Lee,“FuzzyLogic in controlsystems:fuzzylogic controller,” IEEETrans.onSystems,Man,andCybernetics, vol. 20,pp.404–435,1990.

9. E. H. Ruspini, “Fuzzy Logic in the FLAKEY Robot,” in IIZUKA’90, Proceedingsof theInternational Conferenceon FuzzyLogic and Neural Networks,Iizuka/Japan, pp. 202 –206,20.7.- 23.7.1990.

10. Y. Maeda,M. Tanabe,M. Zuta,andT. Takagi,“HierarchicalControlfor AutonomousMobileRobotswith Behavior-DecisionFuzzyAlgorithm,” in Proceddingsof theIEEEInternationalConferenceon RoboticsandAutomation,Nice/France, pp.117– 122,12.5.- 14.5.1992.

11. A. Saffiotti, E. H. Ruspini,andK. Konolige,“Blending Reacitivity andGoal-Directednessin a Fuzzy Controller,” in SecondIEEE International Conferenceon FuzzySystems,SanFrancisco, pp.134– 139,28.3.- 1.4.1993.

12. Y. Watanabeand F. G. Pin, “Sensor-BasedNavigation of a Mobile Robot Using Auto-matically ConstructedFuzzy Rules,” in International Conferenceon AdvancedRobotics,Tokyo/Japan, pp.81– 87,1.11.- 2. 11.1993.

13. P. Reignier, “Fuzzy logic techniquesfor mobile robot obstacleavoidance,” RoboticsandAutonomousSystems, vol. 12,pp.143– 153,1994.

14. E. TunstelandM. Jamshidi,“Fuzzy Logic andBehavior ControlStrategy for AutonomousMobile RobotMapping,” in FUZZ-IEEEWorld Congresson ComputionalIntelligence, Or-lando, pp.514– 517,26.6.- 2.7.1994.

15. H. Surmann,J. Huser, andJ. Wehking,“Path planningfor a fuzzy controlledautonomousmobile robot,” in Fifth International Conferenceon FuzzySystems,FUZZ-IEEE 96, NewOrleans,USA, pp.1660– 1665,Sept.1996.

16. A. Saffiotti, E. H. Ruspini,andK. Konolige,A MultivaluedLogic Approach to IntegratingPlanningand Control. Menlo Park, CA 94025,USA: Artificial IntelligenceCenter, SRIInternational,TechnicalNoteNo. 533,4/1994.

17. A. Saffiotti, “SomeNoteson theIntegrationof PlanningandReactivity in AutonomousMo-bile Robots,” in Procs.of the AAAI SpringSymposiumon Foundationsof AutomaticPlan-ning, (Stanford,CA), pp.122– 126,1993.

18. E.vonPuttkamer, “SensorintegrationzurGeometrischenWeltmodellierung,” it+ti , vol. 1/94,pp.34– 38,1994.

19. V. Gorrini andH. Bersini, “RecurrentFuzzySystems,” in FUZZ-IEEEWorld CongressonComputionalIntelligence, Orlando, pp.193– 198,26.6.- 2.7.1994.

20. T. InfraLogic,Fuzzy-CDevelopmentSystemUser’s Manual. TogaiInfraLogic,1990.21. S. Thrun, “A LivelongPerspective for Mobile RobotControl,” IntelligentRobotsandSys-

tems,Munich, September12-16, vol. 1, pp.23–30,1994.22. G. Dudek,“Roboticexplorationasgraphconstruction,” IEEETransactionsonRoboticsand

Automation, vol. 7, no.6, pp.859–865,1991.23. H. Surmann,J.Huser, andL. Peters,“A fuzzysystemfor indoormobilerobotnavigation,” in

FourthIEEEInternationalConferenceonFuzzySystems,FUZZ-IEEE95,Yokohama,Japan,pp.83–88,20–24Mar. 1995.Distinguishedwith RobotIntelligenceAward.

24. H. Surmann,K. Heesche,M. Hoh, K. Goser, andR. Rudolf, “Entwicklungsumgebung fürFuzzy-Controllermit neuronalerKomponente,” VDE-Fachtagung: “TechnischeAnwendun-genvonFuzzy-Systemen”,Dortmund, pp.288–297,12–13Nov. 1992.

25. H. Surmann,A. P. Ungering,T. Kettner, andK. Goser, “What kind of hardwareis necessaryfor afuzzyrulebasedsystem,” in FUZZ-IEEEWorldCongressonComputionalIntelligence,Orlando, pp.274– 278,26.6.- 2.7.1994.