机器人原文.docx
《机器人原文.docx》由会员分享,可在线阅读,更多相关《机器人原文.docx(14页珍藏版)》请在冰豆网上搜索。
机器人原文
WHISKERBASEDMOBILEROBOTNAVIGATION
DavidJungandAlexanderZelinsky
IntelligentRoboticsLaboratory
ComputerScienceDepartment,UniversityofWollongong
Wollongong,NSW2500,Australia.
djung@cs.uow.edu.au
1.INTRODUCTION
Oneoftherequirementsofourresearch,andacommonrequirementforanautonomousmobilerobots,istheabilitytonavigatecompetentlyinanunstructuredenvironment.Thisproblemhasreceivedmuchresearchattentionintheroboticscommunity.Amajordifficultyisthattherearetwocompetingaspectstonavigation.Arobotmustbeabletoreactveryquicklytoitsenvironmenttopreventcollisionswithstaticanddynamicobstacles.Also,inordertoachieveusefulpurposivenavigationarobotmustuseamap-likerepresentationofitsenvironment.Thisimpliessomesortoflimitedcognition(asdefinedby[McFarland91]).Theformerimpliesminimalprocessingtomeetthereal-timerequirements,whilethelaterrequiresmoreprocessingthancantypicallybeachievedforreal-timeresponse.Thebasiccontroldesignapproachescanbebroadlydividedintofourtypes.Thesearedefinedby[Mataric92a]andwebrieflyreiteratethemandtheirshortcomingshere.Thepurelyreactiveapproachesuseamappingfromsensorsetstoassociatedactions;asetofrules[Brooks87].Theplanner-basedstrategiesoriginatedwiththesymbolicAIcommunityandemployasense-plan-actcycle.Theplanstageusescognitivetechniquestoreasonaboutasymbolicworldmodel.
Therealsoexisthybridsystemswhichemployreactivecomponentsbeneathplanner-basedsystemstoprovidethebenefitsofboth.Anotherapproachisbehaviour-based,whichusesasetofinteractingdistributedconcurrentbehaviours,eachofwhichmayincorporatememoryandlearningofenvironmentrepresentations.
Figure1-YamabicoRobotwithWhiskers
Thepurelyreactiveapproachachievesrobustrealtimeperformance,butthetasksthatcanbeachievedarelimitedbecauseofthelackofanycognition.Theplanningapproachsuffersfromanumberofproblemsincludingslowinteractionwiththeenvironmentduetoslowprocessing,theframeproblemandthesymbolgroundingproblem.Thehybridapproachesattempttomarrytwoincompatiblephilosophiesandstillsufferfrommanyoftheproblemsoftheplanningapproaches.Formoredepthontheseproblemsreferto[Pfeifer95].Thebehaviour-basedapproachhasthepotentialofreal-timeresponseandcognitivisticprocessinginauniformmanner.
Aspartofourongoingresearchintocooperativeheterogeneousmulti-robotsystems,describedbelow,werequirepurposivenavigationandhighspeedwallfollowingatcloserange,upto1m/secat30mm.Thisimmediatelyprecludestheuseofpurelyreactiveorplannerbasedarchitectures.Henceweareadoptingthebehaviour-basedapproach.Alsowehavedevelopedanovelwhiskersensortailoredtoourtask.
Theremainderofthispaperdescribesboththewhiskersensorandthenavigationsystembasedonthewhiskerwehavedeveloped.Thefollowingsectiondescribesourcooperativemulti-robotresearchprojectforwhichthewhiskerbasednavigationwasdeveloped.Section2describesourwhiskersensorsandhowtheyrelatetoexistingwhiskers.Thesectionalsodescribeshowthewhiskersareusedtodetect‘natural’landmarksinthelaboratoryenvironment.Section3detailstechniquesforwhiskerbasedwallfollowingandfinallythelasttwosectionsdiscussthegeneralarchitectureandspecificorganisationofbehaviourunitsusedtoaccomplishwhiskerbasednavigation.
1.1TheContext
Thissectionbrieflydescribestherobotsandtheresearchintocooperationbetweenmultiplerobotsforwhichthewhiskerbasednavigationwasdeveloped.Theaimoftheresearchistoassesstheeffectonperformanceofchangingsomeparametersofthecooperationtechniqueemployed.Thecontextoftheassessmentisaspecificcooperativetaskperformedbyour‘Yamabico’autonomousmobilerobots[Yuta91],oneofwhichispicturedinFigure1.ThetaskwehavechosenisfortwoYamabicorobotstocleanourlaboratoryfloor.Eachhasdifferenttoolssuchthatneithercanaccomplishthetaskalone.Onerobotisequippedwithavacuumcleanerandtheotherwithasweepforsweepingclosetowalls,asisexplainedbelow.Formoreinformationonthecooperationaspectrefertoourpreviouspublication[Jung96].
Therobotseachhavetwowheelscontrolledusingshaftencoderfeedback,andfourultrasonicrangesensors,onefacingeachcompassdirectionrelativetotherobot.Therobotsalsohavethreeprocessorcardsconnectedviaabus,twoMotorola68000’swhichmanagethemainapplicationprogramandthefourultrasonicsensorsandoneINMOST805forthelocomotioncontrol.
Oneoftherobots,thevacuum,hasavacuumcleanercapableofbeingcontrolledviasoftware.It’staskistovacuumpilesoflitterfromthelaboratoryfloor.Itcannotvacuumclosetowallsorfurniture.ItalsohasaCCDcameramountedontopwhichsendsvideosignalsviaavideotransmittertoourvisionsystem.
TheMEPtrackingvisionsystemwasdevelopedbyFujitsuR&DJapan.Itisdesignedforreal-timetrackingofmultipleobjectsinblackandwhiteframesofNTSCvideo.Thehardwareconsistsofatrackingmodulewhichcantrackupto72objectsatframerate(30Hz).AprocessorcardrunningtheVxWorksoperatingsystemexecutestheapplicationprogramandcontrolsthevisioncardviaaVME-bus.Thetrackingofobjectsisbasedonsimple88or1616templatecomparisonusingcrosscorrelationwhichisbasedontheblock-matchingsystemdevelopedbyInoue’sLaboratoryatUniversityofTokyo[Inoue85].
ThevisionsystemisconnectedviaethernettoourhostUNIXworkstationwhichhasaserialconnectiontoaradiomodem.Therobotalsohasaradiomodemandusingthenetworkingsoftwarewedevelopedthevisualmotorpathiscomplete.Thevisionsystemallowsitto‘see’pilesoflitter,butnotfineparticlesscatteredoverthefloor.
Figure2-Thevisionsystemsegmentingthe
imageintoregionsoffloorcarpetandobstacle.
Thevisionsystem’smatchingcapabilityhasalsobeenusedtoimplementarobustcollisionavoidancebehaviour[Zelinsky95b]forthevacuum.Figure2showstheviewfromtherobot,wheretheimagehasbeendividedintoacoursegrid.Eachsquareinthegridisclassifiedbasedonhowwellitmatchesagainstatemplateofthelaboratorycarpet.Thereisaseparatecarpettemplateforeachrowinthegridtoallowforthetexturescalingasthecarpetbecomesfartherawaytowardsthetopoftheimage.Foreachelementofthegrid,thevisionsystemgivesadistortionvaluewhichindicatestheconfidenceofthematchbetweenthecarpettemplateandtheimage.Nextthemainprogramthresholdsthedistortiontoclassifythegridelementaseithercarpetornon-carpet.Thevisualbehaviourforcollisionavoidancemapstheimagespaceintothefloorandphysicallydirectstherobottoavoidareasthatwerenotmatchedascarpet.Althoughthismatchingprocesscontainsnoise,sincetheclassificationisre-evaluatedatframeratearobustcollisionavoidancebehaviourresults.
Theotherrobot,thesweep,hasabrushtoolthatisdraggedoverthefloortosweepdistributedlitterintolargerpilesforthevacuumtopickup.Therobotisrequiredtosweepveryclosetowallswerethevacuumcannotreach.Thiswillrequirebasiccompetenceatnavigatingaroundourclutteredlaboratory,includingobstacleavoidance,andtheabilitytofollowalongwallscloselyathighspeed.Thiscannotbeachievedusingexistingwhiskers,asisshowninthesectiontofollow.
2.WHISKERS
Torealisethesweepingtasktherobotrequiredsensorstoreliablyfollowwallsatcloserangeandhighspeedandsimultaneouslyavoidobstacles.Therobotwasinitiallyequippedwithfourultrasonicrangesensorsasshownbelow.
Figure3-Schematiclayoutofsensors
Theultrasonicrangesensorsprovednottoberobustforwallfollowingbecausetheyoftengiveincorrectdistancereadingsforspecularsurfacesthatareinclinedwithrespecttotherobot.Inadditionthesensorswehavecannotmeasuredistancesunderthe5cmwerequireforclosefollowing.Infraredsensorswereconsidered,butaretypicallydifficulttomakerobustforawidevarietyofsurfacematerialtypesandcolours.
Nextweconsideredcontactsensors,suchaswhiskers.Mostcontactsensorsfortactilesensinginuseonrobotstodatecanbedividedintotwotypes.Thosewhichmeasurestrainandthosethatmeasuredeformation.Mosttactilesensorsthatmeasurethestrainareusedinsystemswithmanipulators.Forareviewsee[Nicholls89]and[Fearing90].Activewhiskersensorsthatmeasurestrainforrecoveringobjectshapehavealsobeenreported,forexampleby[Russell85],buttheseareprohibitivelyexpensivefortheuseofmanyonasmallinexpensivemobilerobotandprovidemoreinformationthanwerequire.Ofthetactilesensorsthatmeasuredeformation,onecommontypeistheskinlikesensors,thattypicallyarecomprisedofanarrayofsimplesensorsformeasuringdeformationataparticularpoint.Theseareusuallyexpensiveandmeasuredeformationsonlyoverasmallrange.Thephysicalsensorsusedmaybeoptical,magnetic,resistive,pneumaticorevenultrasonic,forexample[Russell90].Anothercommontypeofsensorformeasuringdeformationisthemechanicalwhisker.Thistypeofwhisker,asisinuseonmobilerobotstodate,providesonlybinarydeformationinformation;thatis,contact/non-contactinformation(referto[Everett88],[Russell90]forexamples).Acommonandcheapdesigninvolvesmountingpianowiresuchthatdeformationcontactsitwithaconductingplatehencemakingacircuitasshownbelow.
Figure4-Abinarywhisker1
Forwallfollowingtheinformationrequiredforcontrolof