aHumanoidRobot
AndrewG.Brooks
RoboticLifeGroupMITMediaLaboratory
77MassachusettsAvenue,Cambridge,MA,USA
zoz@media.mit.edu
RonaldC.Arkin
MobileRobotLaboratoryandGVUCenterCollegeofComputing,GeorgiaInstituteofTechnology
Atlanta,GA,USAarkin@cc.gatech.edu
July26,2006
Abstract
Thisresearchdetailstheapplicationofnon-verbalcommunicationdisplaybehaviorstoanautonomoushumanoidrobot,includingtheuseofproxemics,whichtodatehasbeenseldomexploredinthefieldofhuman-robotinteraction.Inordertoallowtherobottocommunicateinformationnon-verballywhilesimultaneouslyfulfillingitsexistinginstrumentalbe-havior,a“behavioraloverlay”modelthatencodesthisdataontotherobot’spre-existingmotorex-pressionisdevelopedandpresented.Thestateoftherobot’ssystemofinternalemotionsandmoti-vationaldrivesisusedastheprincipaldatasourcefornon-verbalexpression,butinorderfortherobottodisplaythisinformationinanaturalandnu-ancedfashion,anadditionalpara-emotionalframe-workhasbeendevelopedtosupporttheindividual-ityoftherobot’sinterpersonalrelationshipswithhu-mansandoftherobotitself.AnimplementationontheSonyQRIOisdescribedwhichoverlaysQRIO’sexistingEGOarchitectureandsituatedschema-basedbehaviorswithamechanismforcommunicatingthisframeworkthroughmodalitiesthatencompasspos-ture,gestureandthemanagementofinterpersonaldistance.
1Introduction
Whenhumansinteractwithotherhumans,theyuseavarietyofimplicitmechanismstoshareinforma-tionabouttheirownstateandthestateofthein-teraction.Expressedoverthechannelofthephys-icalbody,thesemechanismsarecollectivelyknown
1
asnon-verbalcommunicationor“bodylanguage”.Ithasnotbeenproventhathumansrespondinpreciselythesamewaytothebodylanguageofahumanoidrobotastheydotothatofahuman.Norhavethespecificrequirementsthattherobotmustmeetinor-dertoensuresucharesponsebeenempiricallyestab-lished.Ithas,however,beenshownthathumanswillapplyasocialmodeltoasociablerobot(Breazeal,C.2003),andwillinmanycasesapproachinterac-tionswithelectronicmediaholdingasetofprecon-ceivedexpectationsbasedontheirexperiencesofin-teractionswithotherhumans(Reeves,B.andNass,C.1996).Ifthesesocialequivalencesextendtotheinterpretationofhuman-likebodylanguagedisplayedbyarobot,itislikelythattherewillbecorrespond-ingbenefitsassociatedwithenablingrobotstosuc-cessfullycommunicateinthisfashion.ForarobotsuchastheSonyQRIO(Figure1)whoseprincipalfunctionisinteractionwithhumans,weidentifythreesuchpotentialbenefits.
Firstisthepracticalbenefitofincreasingthedatabandwidthavailableforthe“situationalawareness”ofthehuman,bytransmittingmoreinformationwithoutaddingadditionalloadtoexistingcommuni-cationmechanisms.Ifitisassumedthatitisbenefi-cialforthehumantobeawareoftheinternalstateoftherobot,yettherearecasesinwhichitisdetrimen-talfortherobottointerruptotheractivities(e.g.dia-log)inordertoconveythisinformation,anadditionalsimultaneousdatachanneliscalledfor.Non-verbalcommunicationisanexampleofsuchachannel,andprovidedthatthecuesareimplementedaccordingtoculturalnormsandareconvincinglyexpressiblebytherobot,addstheadvantageofrequiringnoaddi-tionaltrainingforthehumantointerpret.
Thesecondbenefitistheforestallingofmiscom-municationwithintheexpandedavailabledataband-width.Theproblemraisedifhumansdoindeedhaveautomaticandunconsciousexpectationsofreceivingstateinformationthroughbodilysignals,isthatcer-tainstatesarerepresentedbynullsignals,andhu-mansinteractingwitharobotthatdoesnotcommu-nicatenon-verbally(orwhichdoessointermittentlyorineffectively)maymisconstruelackofcommunica-tionasdeliberatecommunicationofsuchastate.Forexample,failingtorespondtopersonalverbalcom-municationswithattentivesignals,suchaseyecon-tact,cancommunicatecoldnessorindifference.Ifa
humanoidrobotisequippedwiththosesortsofemo-Figure1:SonyQRIO,anautonomoushumanoidtions,itisimperativethatwetrytoensurethatsuchrobotdesignedforentertainmentandinteraction“falsepositive”communicationsareavoidedtothewithhumans,shownhereinitsstandardposture.fullestextentpossible.
Thethirdpotentialbenefitisanincreasedproba-bilitythathumanswillbeabletoformbondswiththerobotthatareanalogoustothoseformedwithotherhumans—forexample,affectionandtrust.Webe-lievethatthedevelopmentofsuchrelationsrequiresthattherobotappear“natural”—thatitsactionscanbeseenasplausibleinthecontextoftheinter-nalandexternalsituationsinwhichtheyoccur.Inotherwords,ifapersoncollaboratingwiththerobotcan“ataglance”gainaperspectiveofnotjustwhattherobotisdoingbutwhyitisdoingit,andwhatitislikelytodonext,wethinkthatheorshewillbemorelikelytoapplyemotionallysignificantmodelstotherobot.AprincipaltheoryconcerninghowhumanscometobesoskilledatmodelingothermindsisSim-ulationTheory,whichstatesthathumansmodelthemotivationsandgoalsofanobservedagentbyusingtheirowncognitivestructurestomentallysimulatethesituationoftheobservee(Davies,M.andStone,T.1995,Gordon,R.1986,Heal,J.2003).Thissug-geststhatitislikelythatthemoretheobservablebehavioroftherobotdisplaysitsinternalstatebyreferencingthebehaviorsthehumanhasbeencondi-tionedtorecognize—themoreit“actslike”ahumaninits“displaybehaviors”—themoreaccuratethehuman’smentalsimulationoftherobotcanbecome.Withthesebenefitsinmindasultimategoals,weherebyreportonactivitiestowardsthemoreimme-diategoalofrealizingthedisplaybehaviorsthem-selves,underthreespecificconstraints.First,suchdisplaysshouldnotrestrictthesuccessfulexecutionof“instrumentalbehaviors”,thetaskstherobotisprimarilyrequiredtoperform.Second,theapplica-tionofbodylanguageshouldbetightlycontrolledto
2
avoidconfusingthehuman—itmustbeexpressedwhenappropriate,andsuppressedwhennot.Third,non-verbalcommunicationinhumansissubtleandcomplex;therobotmustsimilarlybeabletousethetechniquetorepresentarichmeshingofemotions,motivationsandmemories.Tosatisfytheserequire-ments,wehavedevelopedtheconceptofbehavioral“overlays”forincorporatingnon-verbalcommunica-tiondisplaysintopre-existingrobotbehaviors.First,overlaysprovideapracticalmechanismformodifyingtherobot’spre-existingactivities“on-the-fly”withexpressiveinformationratherthanrequiringthede-signofspecificnewactivitiestoincorporateit.Sec-ond,overlayspermitthepresenceorabsenceofbodylanguage,andthedegreetowhichitisexpressed,tobecontrolledindependentoftheunderlyingactivity.Third,theoverlaysystemcanbedrivenbyanarbi-trarilydetailedmodelofthesedrivingforceswithoutforcingthismodeltobedirectlyprogrammedintoeveryunderlyingbehavior,allowingevensimpleac-tivitiestobecomemorenuancedandengaging.
Abriefsummaryofthecontributionsofthisre-searchfollows:
1.Thisworkbroadensandreappraisestheuseofbodilyexpressivenessinhumanoidrobots,par-ticularlyintheformofproxemics,whichhashithertobeenonlyminimallyconsideredduetosafetyconsiderationsandtherelativescarcityofmobilehumanoidplatforms.2.Thisworkintroducestheconceptofabehav-ioraloverlayfornon-verbalcommunicationthatbothencodesstateinformationintothephysicaloutputofordinarybehaviorswithoutrequiring
modificationtothebehaviorsthemselves,andin-creasesnon-verbalbandwidthbyinjectingaddi-tionalcommunicativebehaviorsintheabsenceofphysicalresourceconflicts.
3.Thispaperfurtherdevelopsthebehavioralcom-municationsoverlayconceptintoageneralmodel
suitableforapplicationtootherroboticplat-2.1Proxemicsformsandinformationsources.
Hall,pioneerofthefieldofproxemics,identifieda
4.Incontrastwithmuchpriorwork,theresearchnumberoffactorsthatcouldbeusedtoanalyzedescribedhereprovidesmoredepthtotheinfor-theusageofinterpersonalspaceinhuman-humanmationthatiscommunicatednon-verbally,giv-interactions(Hall,E.T.1966).Statedescriptorsin-ingtherobotthecapabilityofpresentingitsin-cludethepotentialfortheparticipantstotouch,smellternalstateasinterpretedviaitsownindividual-andfeelthebodyheatofoneanother,andthevisualityandinterpersonalmemoryratherthansimplyappearanceoneanother’sfaceataparticulardistanceaninstantaneousemotionalsnapshot.(focus,distortion,dominationofvisualfield).The5.Similarly,whileexpressivetechniquessuchasfa-reactionsofindividualstoparticularproxemicsitu-cialfeatureposesarenowfrequentlyusedinationsweredocumentedaccordingtovariouscodes,robotstocommunicateinternalstateandengagemonitoringaspectssuchastheamountandtypeofthehuman,thisworkmakesprogresstowardsthevisualandphysicalcontact,andwhetherornottheuseofbodilyexpressioninagoal-directedfash-bodypostureofthesubjectswasencouraging(“so-ciopetal”)ordiscouraging(“sociofugal”)ofsuchcon-ion.
tact.
6.Finally,thisworkpresentsafunctioningimple-Aninformalclassificationwasusedtodividethe
mentationofabehavioraloverlaysystemonacontinuousspaceofinterpersonaldistanceintofourrealrobot,includingthedesignofdatastruc-generalzonesaccordingtothesestatedescriptors.Inturestorepresenttherobot’sindividualrespon-orderofincreasingdistance,theseare“Intimate”,sivenesstospecifichumansandtoitsowninter-“Personal”,“Socio-Consultive”and“Public”.Hu-nalmodel.manusageofthesespacesinvariousrelationshipsand
situationshasbeenobservedandsummarized(Weitz,S.1974),andcanbeusedtoinformtheconstruc-2ProxemicsandBodyLan-tionofaroboticsystemthatfollowssimilarguidelines
(subjecttovariationsinculturalnorms).guage
Spatialseparationmanagementthereforehasprac-Behavioralresearchershavecomprehensivelyenumer-ticaleffectsintermsofthepotentialforsensingandatedandcategorizedvariousformsofnon-verbalphysicalcontact,andemotionaleffectsintermsofthecommunicationinhumansandanimals.Inconsider-comfortoftheparticipantswithaparticularspatialingnon-verbalcommunicationforahumanoidrobot,arrangement.Whatconstitutesanappropriatear-wehaveprimarilyfocusedonthemanagementofspa-rangementdependsonthenatureoftheinteractiontialrelationshipsandpersonalspace(proxemics)and(whatkindsofsensingandcontactarenecessary,andonbodilyposturesandmovementsthatconveymean-whatkindsofemotionalstatesaredesiredforit),theing(kinesics).Thelatterclasswillbemorelooselyrelationshipbetweentheparticipants(anappropri-referredtoas“bodylanguage”,tounderscorethefactatedistanceforamarriedcouplemaybedifferentthatwhilemanykinesicgesturescanconveymeaningthanthatbetweenbusinessassociatesengagedintheintheirownright,perhapsthemajorityofkinesicsameactivity),andthecurrentemotionalstatesofcontributionstonon-verbalcommunicationoccursintheparticipants(theprecedingfactorsbeingequal,aparalinguisticcapacity,asanenhancementofcon-theshapeandsizeofanindividual’sidealpersonalcurrentverbaldialog(Dittmann,A.1978).Theuse“envelope”canexhibitsignificantvariationbasedonofthistermisnotintended,however,toimplythathisorherfeelingsatthetime,asshowninFigure2).theseposturesandmovementsformatruelanguageToensurethatarobotdisplaysappropriateusagewithdiscreterulesandgrammars;butasMachotkaofandrespectforpersonalspace,andtoallowittoandSpiegelpointout,theyconveycodedmessagestakeactionstomanipulateitinwaysthatthehuman
3
thathumanscaninterpret(Machotka,P.andSpiegel,J.1982).Takentogether,proxemicsandbodylan-guagecanreflectsomeorallofthetypeofinteraction,therelationsbetweenparticipants,theinternalstatesoftheparticipantsandthestateoftheinteraction.
Figure2:Illustrationofhowanindividual’s‘per-sonal’spacezonemayvaryinsizeandshapeaccord-ingtoemotion.Duringfear,thespaceanindivid-ualconsidershisownmightexpand,withgreaterex-pansionoccurringtohisrearasheavoidspotentialthreatsthathecannotsee.Duringanger,thespaceanindividualconsidersherownmightexpandtoagreaterextenttoherfrontasshedirectshercon-frontationalattentiontoknownpresences.ProxemicFactorKinestheticpotential(armsonly)
Kinestheticpotential(armsplustorso)Minimumfacerecog-nitiondistance
Human60–75cm90–120cm<5cm
QRIO20cm25–35cm20cm
Figure3:QRIO’sproxemiczonesinthisimplemen-tation,selectedasabalancebetweenthezonesforanadulthumanandthosecomputedusingQRIO’srel-evantproxemicfactors.Thedemarcationdistancesbetweenzonesrepresentthemidpointofafuzzythresholdfunction,ratherthana‘hard’cutoff.investigations.However,somerelatedworkexists.Themobilerobot‘Chaser’byYamasakiandAnzaifocusedononeofthepracticaleffectsofinterpersonaldistancebyattemptingtosituateitselfatadistancefromthehumanthatwasidealforsensoroperation,inthiscasethecollectionofspeechaudio(Yamasaki,N.andAnzai,Y.1996).Thisworkdemonstratedthatawarenessofpersonaldistanceconsiderationscouldbeactedupontoimprovespeechrecognitionperfor-mance.
Inasimilarvein,Kandaetal.performedahuman-robotinteractionfieldstudyinwhichtherobotdis-tinguishedconcurrentlypresenthumansaseitherparticipantsorobserversbasedontheirproxemicdistance(Kanda,T.,Hirano,T.,Eaton,D.andIshig-uro,H.2004).Asinglefixeddistancethresholdwasusedfortheclassification,however,anditwasleftuptothehumanparticipantstomaintaintheappropri-ateproxemics.
LikhachevandArkinexploredthenotionof“com-fortzones”foramobilerobot,usingattachmentthe-orytoinformanemotionalmodelthatrelatedtherobot’scomforttoitsspatialdistancefromanobjectofattachment(Likhachev,M.andArkin,R.C.2000).Theresultsofthisworkshowedthattherobot’sex-plorationbehaviorvariedaccordingtoitslevelofcomfort;whilethepublishedworkdidnotdealwithhuman-robotinteractiondirectly,usefulHRIscenar-ioscouldbeenvisagedforcasesinwhichtheobjectofattachmentwasahuman.
Morerecently,therehavebeeninvestigationsintomodifyingthespatialbehaviorofnon-humanoidmo-bilerobotsinordertomakepeoplefeelmoreateasewiththerobot.Smithinvestigatedself-adaptationof4
Table1:AcomparisonofselectproxemicfactorsthatdifferbetweenadulthumansandQRIO(equippedwithstandardoptics).
canunderstandandinferfromthemtheunderlyingreasoning,requiresconsiderationofallofthesefac-tors.Inaddition,thesizeoftherobot(whichisnotlimitedtotherangefixedbyhumanbiology)mayhavetobetakenintoaccountwhenconsideringtheproxemicsthatahumanmightbelikelytofindnat-uralorcomfortable.SeeTable1foracomparisonofseveralproxemicfactorsinthecaseofadulthumansandQRIO,andFigure3forthegeneralproxemiczonesthatwereselectedforQRIO.
Therehasbeenlittleexplorationoftheuseofprox-emicsinhuman-robotinteractiontodate.Therea-sonsforthisareperhapsmostlypragmaticinna-ture.Roboticmanipulators,includinghumanoidup-pertorsos,canbedangeroustohumansandinmostcasesarenotrecommendedforinteractionswithindistancesatwhichphysicalcontactispossible.Inaddition,humanoidrobotswithfullmobilityarestillrelativelyrare,andthosewithleggedlocomotion(completehumanoids)evenmoreso,precludingsuch
thebehaviorofaninteractivemobilerobot,includingitsspatialseparationfromthehuman,basedonitsassessmentofthehuman’scomfort(Smith,C.2005).Inthiscasethegoalwasfortherobottoautomati-callylearnthepersonalspacepreferencesofindivid-ualhumans,ratherthantousespatialseparationasageneralformofnon-verbalcommunication.Howeversomeoftheinformalhumansubjectexperimentsre-portedareinstructiveastothevalueoftakingprox-emicsintoaccountinHRI,particularlythecaseinwhichmisrecognitionofahumandiscomfortresponseasacomfortresponseleadstoapositivefeedbackloopthatresultsindistressingbehavioronthepartoftherobot.
Similareffortshaveusedconsiderationsofhu-mans’personalspacetoaffectrobotnavigation.NakauchiandSimmonsdevelopedarobotwhosegoalwastoqueueuptoregisterforaconferencealongwithhumanparticipants(Nakauchi,Y.andSimmons,R.2000).Therobotthusneededtodeterminehowtomovetopositionsthatappropriatelymatchedhumanqueuingbehavior.Althausetal.describeasystemdevelopedtoallowarobottoapproachagroupofpeopleengagedinadiscussion,enterthegroupbyassumingaspatiallyappropriateposition,andthenleaveandcontinueitsnavigation(Althaus,P.,Ishig-uro,H.,Kanda,T.,Miyashita,T.andChristensen,H.I.2004).ChristensenandPacchierottiuseprox-emicstoinformacontrolstrategyfortheavoidancebehaviorexhibitedbyamobilerobotwhenforcedbyaconstrainingpassagewaytonavigateincloseprox-imitytohumans(Christensen,H.I.andPacchierotti,E.2005).Pacchierottietal.thenreportpositiveresultsfromapilotuserstudy,inwhichsubjectspre-ferredtheconditioninwhichtherobotmovedfastestandsignaledthemostdeferencetothehumans(bymovingoutofthewayearliestandkeepingthegreat-estdistanceaway),thoughthesubjectswereallfa-miliarandcomfortablewithrobots(Pacchierotti,E.,Christensen,H.I.andJensfelt,P.2005).Whilein-formedbyproxemicstheory,theseeffortsfocusoncontroltechniquesforapplyingsocialappropriate-nesstonavigationactivity,ratherthanusingutilizingproxemicbehavioraspartofanon-verbalcommuni-cationsuite.
Anotherrecentexperiment,byteBoekhorstetal.,gavesomeconsiderationtopotentialeffectsofthedistancebetweenchildrenandanon-humanoidrobotonthechildren’sattentiontoa“passthepar-cel”game(teBoekhorst,R.,Walters,M.,Koay,K.L.,Dautenhahn,K.andNehaniv,C.2005).Nosignifi-5
canteffectswererecorded,howevertheauthorsad-mitthatthedataanalysiswascomplicatedbyvio-lationsoftheassumptionsunderlyingthestatisticaltestsandthereforewebelievetheseresultsshouldnotbeconsideredconclusive.Datafromthesameseriesofexperimentswasusedtopointoutthatthechil-dren’sinitialapproachdistancesweresociallyappro-priateaccordingtohumanproxemicstheory(Walters,M.L.,Dautenhahn,K.,Koay,K.L.,Kaouri,C.,teBoekhorst,R.,Nehaniv,C.L.,Werry,I.andLee,D.2005a).Afollow-upexperimentwasconductedusingthesamenon-humanoidrobottoinvestigatetheapproachdistancesthatadultspreferredwheninteractingwiththerobot(Walters,M.L.,Dauten-hahn,K.,teBoekhorst,R.,Koay,K.L.,Kaouri,C.,Woods,S.,Nehaniv,C.L.,Lee,D.andWerry,I.2005b).Amajority,60%,positionedthemselvesatadistancecompatiblewithhumanproxemicstheory,whereastheremainder,asignificantminorityof40%,assumedpositionssignificantlycloser.Whiletheseresultsaregenerallyencouragingintheirempiricalsupportofthevalidityofhuman-humanproxemictheorytohuman-robotinteractions,cautionshouldbeobservedinextrapolatingtheresultsineitherdi-rectionduetothenon-humanoidappearanceoftherobot.
Morerecentlythantheinitialsubmissionofthispa-per,therehasbeeninterestshownintheuseofprox-emicsandnon-verbalcommunicationbytheroboticsearchandrescuecommunity.InaconferenceposterBethelandMurphyproposedasetofguidelinesforaf-fectexpressionbyappearance-constrained(i.e.non-humanoid)rescuerobotsbasedontheirproxemiczonewithrespecttoahuman(Bethel,C.L.andMur-phy,R.R.2006).Thegoalofthisworkwasonceagaintoimprovethecomfortlevelofhumansthroughso-ciallyawarebehavior.
Whilenotinvolvingrobotsperse,somevirtualen-vironment(VE)researchershaveexaminedreactionstoproxemicconsiderationsbetweenhumansandhu-manoidcharacterswithinimmersiveVEs.Bailensonetal.,forexample,demonstratedthatpeopleexhib-itedsimilarpersonalspatialbehaviortowardsvirtualhumansastheywouldtowardsrealhumans,andthiseffectwasincreasedthemorethevirtualhumanwasbelievedtobetheavatarofarealhumanratherthananagentcontrolledbythecomputer(Bailenson,J.N.,Blascovich,J.,Beall,A.C.andLoomis,J.M.2003).However,theyalsoencounteredtheinterestingresultthatsubjectsexhibitedmorepronouncedavoidancebehaviorwhenproxemicboundarieswereviolatedby
anagentratherthananavatar;theauthorstheo-rizethatthisisduetothesubjectsattributingmorerationalityandawarenessofsocialspatialbehaviortoahuman-drivenavatarthanacomputer-controlledagent,thus“trusting”thattheavatarwouldnotwalkintotheirvirtualbodies,whereasanagentmightbemorelikelytodoso.ThismaypresentalessonforHRIdesigners:theprecisefactthatanautonomousrobotisknowntobeundercomputercontrolmaymakesociallycommunicativeproxemicawareness(asopposedtosimplecollisionavoidance)particularlyimportantforrobotsintendedtooperateincloseproximitywithhumans.
thebroadrangeofexpressionthatisdetailedwithintheseclassifications,inordertoselectthoseappear-ingtohavethemostutilityforroboticapplications.Forexample,classifiedwithinthesetaxonomiesarebodilymotionsthatcancommunicateexplicitsym-bolicconcepts,deicticspatialreferences(e.g.point-ing),emotionalstatesanddesires,likesanddislikes,socialstatus,engagementandboredom.Asaresulttherehasbeensignificantongoingroboticsresearchoverlappingwithalloftheareasthusreferenced.Foracomprehensivesurveyofsociallyinteractiveroboticresearchingeneral,incorporatingmanyoftheseas-pects,see(Fong,T.,Nourbakhsh,I.andDauten-hahn,K.2003).
Theuseofemblematicgesturesforcommunicationhasbeenwidelyusedonroboticplatformsandinthefieldofanimatronics;examplesaretheMITMe-diaLab’s‘Leonardo’,anexpressivehumanoidwhichcurrentlyusesemblematicgestureasitsonlyformofsymboliccommunicationandalsoincorporateski-nesicadaptorsintheformofblendednaturalidlemotions(Breazeal,C.,Brooks,A.G.,Gray,J.,Hoff-man,G.,Kidd,C.,Lee,H.,Lieberman,J.,Lock-erd,A.andChilongo,D.2004),andWasedaUniver-sity’sWE-4RII‘emotionexpressionhumanoidrobot’whichwasalsodesignedtoadoptexpressivebodypostures(Zecca,M.,Roccella,S.,Carrozza,M.C.,Cappiello,G.,Cabibihan,J.-J.,Dario,P.,Takanobu,H.,Matsumoto,M.,Miwa,H.,Itoh,K.andTakan-ishi,A.2004).
Comprehensivecommunicativegesturemecha-nismshavealsobeenincorporatedintoanimatedhu-manoidconversationalagentsandVEavatars.KoppandWachsmuthusedahierarchicalkinesicmodeltogeneratecomplexsymbolicgesturesfromgesturephrases,laterinterleavingthemtightlywithconcur-rentspeech(Kopp,S.andWachsmuth,I.2000,Kopp,S.andWachsmuth,I.2002).Guye-Vuillemeetal.providedcollaborativeVEuserswiththemeanstomanuallydisplayavarietyofnon-verbalbodilyex-pressionsontheiravatarsusingafixedpaletteofpotentialactions(Guye-Vuilleme,A.,Capin,T.K.,Pandzic,I.S.,Thalmann,N.M.andThalmann,D.1998).
Similarly,illustratorsandregulatorshavebeenusedtopunctuatespeechandcontrolconversa-tionalturn-takingoninteractiverobotsandani-matedcharacters.AoyamaandShimomuraim-plementedcontingentheadpose(suchasnodding)andautomaticfillerinsertionduringspeechinterac-tionswithSonyQRIO(Aoyama,K.andShimomura,6
2.2BodyLanguage
Bodylanguageisthesetofcommunicativebodymo-tions,orkinesicbehaviors,includingthosethatareareflectionof,orareintendedtohaveaninfluenceon,theproxemicsofaninteraction.Knappidentifiesfivebasiccategories:
1.Emblems,whichhavespecificlinguisticmeaningandarewhatismostcommonlymeantbytheterm‘gestures’;2.Illustrators,whichprovideemphasistoconcur-rentspeech;3.AffectDisplays,morecommonlyknownasfa-cialexpressionsandusedtorepresentemotionalstates;4.Regulators,whichareusedtoinfluenceconver-sationalturn-taking;and5.Adaptors,whicharebehavioralfragmentsthatconveyimplicitinformationwithoutbeingtiedtodialog(Knapp,M.1972).
Dittmannfurthercategorizesbodylanguageintodiscreteandcontinuous(persistent)actions,withdis-creteactionsfurtherpartitionedintocategorical(al-waysperformedinessentiallythesameway)andnon-categorical(Dittmann,A.1978).Bodypostureitselfisconsideredtobeakinesicbehavior,inasmuchasmotionisrequiredtomodifyit,andbecausetheycanbemodulatedbyattitude(Knapp,M.1972).Thekindsofbodylanguagedisplaysthatcanberealizedonaparticularrobotofcoursedependonthemechanicaldesignoftherobotitself,andthesecategorizationsofhumanbodylanguagearenotnec-essarilyofprincipalusefulnesstoHRIdesignersotherthansometimessuggestingimplementationaldetails(e.g.thenecessityofpreciselyaligningillustratorswithspokendialog).However,itisusefultoexamine
H.2005).Extensiveworkonbodylanguageforan-ofbodylanguagetotheinteractionsenvisagedforimatedconversationalagentshasbeenperformedatQRIO,thefollowingaspectswerechosenforspecifictheMITMediaLab,suchasThorisson’simplemen-attention:tationofamultimodaldialogskillmanagementsys-I.Proxemicsandthemanagementofinterper-temonananimatedhumanoidforface-to-facein-sonaldistance,includingspeedoflocomo-teractions(Thorisson,K.R.1996)andCassellandtion;Vilhjalmsson’sworkonallowinghuman-controlled
full-bodyavatarstoexhibitcommunicativereactionsII.Emblematichandandarmgesturesinsup-portoftheabove;tootheravatarsautonomously(Cassell,J.andVilh-jalmsson,H.1999).III.Therotationofthetorsoduringinteraction,whichinhumansreflectsthedesireforin-Robotsthatcommunicateusingfacialexpression
teraction(facingmoresquarelyrepresentsahavealsobecomethesubjectofmuchattention,
sociopetalstance,whereasdisplayinganan-toonumeroustosummarizeherebutbeginningwith
gularoffsetisasociofugalposture)—thuswell-knownexamplessuchasKismetandtheface
knownasthe“sociofugal/sociopetalaxis”;robotsdevelopedbyHara(Breazeal,C.2000,Hara,
F.,Akazawa,H.andKobayashi,H.2001).Inad-IV.Thepostureofthearms,includingcontinu-ditiontocommunicationofemotionalstate,someousmeasures(armsakimbo,defensiverais-oftheserobotshaveusedaffectivefacialexpressioningofthearms,androtationofthefore-withtheaimofmanipulatingthehuman,eitherinarmswhichappearsociopetalwhenrotatedtermsofadesiredemotionalstateasinthecaseofoutwardsandsociofugalwhenrotatedin-Kismetorintermsofincreasingdesiredmotivationwards)anddiscreteposturalstances(e.g.toperformacollaborativetaskasinsubsequentworkarmsfolded);onLeonardo(Brooks,A.G.,Berlin,M.,Gray,J.and
V.Headposeandthemaintenanceofeyecon-Breazeal,C.2005).
tact;
Howevermuchofthisrelatedworkeitherfocuses
directlyonsocialcommunicationthroughbodylan-VI.Illustrators,bothpre-existing(headnod-ding)andnewlyincorporated(attentiveguageasthecentralresearchtopicratherthanthein-torsoleaning).teroperationofnon-verbalcommunicationwithcon-currentinstrumentalbehavior,oronimprovementsSeeFigure15inSection6forexamplesofsomeof
totheinteractionresultingfromdirectlyintegrat-theseposesasdisplayedbyQRIO.ingnon-verbalcommunicationaspartoftheinter-actiondesignprocess.Whenemotionalmodelsareincorporatedtocontrolaspectssuchasaffectivedis-3BehavioralOverlaysplay,theytendtobemodelsdesignedtoprovidea
“snapshot”oftherobot’semotionalstate(forexam-Theconceptofabehavioraloverlayforrobotscanplerepresentedbyanumberofdiscretecategoriesbedescribedasamotor-levelmodifierthataltersthesuchastheEkmanmodel(Ekman,P.andDavid-resultingappearanceofaparticularoutputconfor-son,R.J.1994))suitableforimmediatecommunica-mation(amotion,posture,orcombinationofboth).tionviatherobot’sfacialactuatorsbutwithminimalTheintentionistoprovideasimultaneousdisplayreferencetothecontextoftheinteraction.Howeverofinformation,inthiscasenon-verbalcommunica-recentresearchbyFridlundhasstronglychallengedtion,throughcarefulalterationofthemotorsystemthewidelyacceptednotionthatfacialexpressionsareinsuchawaythattheunderlyingbehavioralactivi-anunconsciousandlargelyculturallyinvariantrep-tiesoftherobotmaycontinueasnormal.Justastheresentationofinternalemotionalstate,arguingin-behaviorschemasthatmakeuptherobot’sbehav-steadthattheyareverydeliberatecommunicationsioralrepertoiretypicallyneednotknowoftheexis-thatareheavilyinfluencedbythecontextinwhichtenceormethodofoperationofoneanother,behav-theyareexpressed(Fridlund,A.1994).Thisisacon-ioraloverlaysshouldbelargelytransparenttothetentionthatmaybeworthkeepinginmindconcern-behaviorschemaresponsibleforthecurrentinstru-ingroboticbodylanguageexpressioningeneral.mentalbehavioratagiventime.PreservationofthisGiventhecapabilitiesofourroboticplatformlevelofmodularitysimplifiestheprocessofaddingor(SonyQRIO)andtherelevanceofthevarioustypeslearningnewbehaviors.
7
Asasimplifiedexample,considerthecaseofasim-ple2-DOFoutputsystem:thetailofaroboticdogsuchasAIBO,whichcanbeangledaroundhorizontalandverticalaxes.Thetailisusedextensivelyfornon-verbalcommunicationbydogs,particularlythroughvariousmodesoftailwagging.FourexamplesofsuchcommunicationsviathetailfromtheAIBOmotorprimitivesdesignspecificationarethefollowing:•TailWaggingFriendly:Amplitudeofwaglarge,heightoftaillow,speedofwagbaselineslowbutrelatedtostrengthofemotion.
•TailWaggingDefensive:Amplitudeofwagsmall,heightoftailhigh,speedofwagbaselinefastbutrelatedtostrengthofemotion.•SubmissivePosture:Incasesoflowdomi-nance/highsubmission,heightoftailverylow(betweenlegs),nowaggingmotion.
•“ImperiousWalking”:Simultaneouswithloco-motionincasesofhighdominance/lowsubmis-sion;amplitudeofwagsmall,heightoftailhigh,speedofwagfast.
Onemethodofimplementingthesecommunicationmodeswouldbetoplacethemwithineachbehavioralschema,designingthesebehaviorswiththeincreasedcomplexityofrespondingtotherelevantemotionalandinstinctmodelsdirectly.Analternativeapproach—behavioraloverlays—istoallowsimplerunder-lyingbehaviorstobeexternallymodifiedtoproduceappropriatedisplayactivity.Considerthebasicwalk-ingbehaviorb0inwhichthedog’stailwagsnaturallyleftandrightatheightφ0withamplitude±θ0andfrequencyω0fromdefaultvaluesforthewalkingstepmotionratherthanthedog’semotionalstate.ThemotorstateMTofthetailundertheseconditionsisthusgivenby:
b0x(t)θ0sin(ω0t)
MT(t)==
b0y(t)φ0Nowconsiderabehavioraloverlayvectorforthe
tailo0=[α,λ,δ]appliedtotheactivebehaviorac-cordingtothemathematicsofamultidimensionaloverlaycoordinationfunctionΩTtoproducethefol-lowingoverlaidtailmotorstateM+T:
αθsin(λωt)00
M+T(t)=ΩT(o0,b0(t))=φ0+δ
λ1,δ0)aswellasothercommunicativewalk
stylesnotspecificallypredefined(e.g.“submissivewalking”:α=0,δ0),withoutanymodifica-tionoftheunderlyingwalkingbehavior.Moreover,thedisplayoutputwillcontinuetoreflectasmuchaspossibletheparametersoftheunderlyingactivity(inthiscasethewalkingmotion)inadditiontotheinternalstateusedtogeneratethecommunicationsoverlay(e.g.dominance/submission,emotion).
HoweverinourAIBOexamplesofar,theoverlaysystemisonlyabletocommunicatethedog’sinternalstateusingthetailwhenitisalreadybeingmovedbyanexistingbehavior(inthiscasewalking).Itmaythereforebenecessarytoaddanadditionalspecialtypeofbehaviorwhosefunctionistokeeptheover-laysystemsuppliedwithmotorinput.Thistypeofbehaviorisdistinguishedbytwocharacteristics:adifferentstimulussetthannormalbehaviors(eithermoreorless,includingnoneatall);anditsoutputistreateddifferentlybytheoverlaysystem(whichmayattimeschoosetoignoreitentirely).Werefertosuchbehaviorsas“idler”behaviors.Inthiscase,consideridlerbehaviorb1whichsimplyattemptstocontinu-ouslywagthetailsomeamountinordertoprovidetheoverlaysystemwithinputtobeaccentuatedorsuppressed:
θ1sin(ω1t)
b1(t)=
φ1Thisbehaviorcompetesforactionselectionwiththeregular“environmental”behaviorsasnormal,andwhenactiveisoverlaidbyΩTinthesamefash-ion.Thustheoverlaysystemwiththeadditionofoneextremelybasicbehaviorisabletoachieveallofthefourtailcommunicationsdisplaysoriginallyspecifiedabove,includingvariationsindegreeandcombina-tion,byappropriateselectionofoverlaycomponentsbasedontherobot’sinternalstate.Foractivebe-haviorbiandoverlayoj,thetailactivityproducedis:
M+T(t)=ΩT(oj,bi(t))
However,theadditionofspecialized“idler”behav-iorsprovidesadditionalopportunitiesformanipula-tionoftherobot’sdisplayactivity,asthesebehaviors
canbedesignedtobeawareofandcommunicatewiththeoverlaysystem—forexample,toenablethetrig-geringofemblematicgestures.Iftherobot’snormalbehaviorsaresubjectatagiventimetothestimu-Forappropriateconstructionofo0,theoverlaysys-lusvector[S],theidlerbehaviorscanbethoughtoftemisnowabletoproduceimperiouswalking(α1,asrespondingtoanexpandedstimulusvector[S,Ψ]
8
themselvesagainstmodificationincasesinwhichin-terferenceislikelytocausefailureofthebehavior(suchasatask,likefinemanipulation,thatwouldsimilarlyrequireintenseconcentrationandsuppres-sionofthenon-verbalcommunicationchannelwhenperformedbyahuman).Careshould,ofcourse,betakentousethisfacilitysparingly,inordertoavoidtheinadvertentsendingofnon-verbalnullsig-nalsduringactivitiesthatshouldnotrequiresuchconcentration,orforwhichtheconsequencesoffail-urearenotexcessivelyundesirable.
SchemasmayalsomakerecommendationstotheFigure4:Thebehavioraloverlaymodel,shownover-layingactiveenvironmentalbehaviorsb1..nandidleroverlaysystemthatassistitinsettingtheenvelope
behaviorsi1..mafteractionselectionhasalreadybeenofpotentialoverlays(suchasreportingthecharac-teristicproxemicsofaninteraction;e.g.whetheraperformed.
speakingbehaviortakesplaceinthecontextofanintimateconversationorapublicaddress).Ingen-whereΨisthevectoroffeedbackstimulifromtheeral,however,knowledgeofandcommunicationwithoverlaysystem.Forinstance,AIBO’stailidlerbe-theoverlaysystemisnotarequirementforexecutionhavioruponreceivingafeedbackstimulusψpmightofabehavior.AsQRIO’sintentionalmechanismisinterruptitswaggingactiontotraceoutapredefinedrefinedtobetterfacilitatehigh-levelbehavioralcon-shapePwiththetailtip:trol,theintentionsystemmayalsocommunicatewith
thebehavioraloverlaysystemdirectly,preservingthe
Px(t)modularityoftheindividualbehaviors.b1(ψp,t)=
Py(t)
Relatedworkofmostrelevancetothisconceptis
Ingeneral,then,letussaythatforacollectionthegeneralbodyofresearchrelatedtomotionpa-ofactive(i.e.havingpassedactionselection)envi-rameterization.TheessentialpurposeofthisclassronmentalbehaviorsBandidlerbehaviorsI,andoftechniquesistodescribebodilymotionsintermsanoverlayvectorO,theoverlaidmotorstateM+ofparametersotherthantheirbasicjoint-angletime
series.Ideally,thenewparametersshouldcaptureisgivenaccordingtothemodel:
essentialqualitiesofthemotion(suchasitsoverallappearance)insuchawaythatthesequalitiescanbeM+=Ω(O,[B(S),I([S,Ψ]))
predictablymodifiedorheldconstantbymodifying
ThismodelisrepresentedgraphicallyinFigure4.orholdingconstanttheappropriateparameters.ThisIntheidealizedmodularcase,theenvironmentalbe-hasadvantagesbothformotiongeneration(classeshaviorsneedneithercommunicatedirectlywithnorofmotionscanberepresentedmorecompactlyaspa-evenbeawareoftheexistenceofthebehavioralover-rameterrangesratherthanclustersofindividualex-laysystem.Forpracticalpurposes,however,acoarseemplars)andmotionrecognition(novelmotionscanlevelofinfluencebytheenvironmentalbehaviorsonbematchedtoknownexamplesbycomparisonofthetheoverlaysystemisrequired,becauseacompleteparametervalues).determinationaprioriofwhetherornotmotormodi-ficationwillinterferewithaparticularactivityisverydifficulttomake.Thislevelofinfluencehasbeenac-countedforinthemodel,andtheinputtoΩfromBandIasshownincorporatesanynecessarycom-municationaboveandbeyondthemotorcommandsthemselves.
Behavioraloverlaysasimplementedintheresearchdescribedinthispaperincludesuchfacilitiesforschemastocommunicatewiththeoverlaysystemwhennecessary.Schemasmay,ifdesired,protect
9
Approachestothistechniquehavebeenappliedtomotiongenerationforanimatedcharacterscandifferintheirprincipalfocus.Onephilosophyin-volvescreatingcomprehensivesetsofbasicmotiontemplatesthatcanthenbeusedtofashionmorecom-plexmotionsbyblendingandmodifyingthemwithasmallersetofbasicparameters,suchasduration,amplitudeanddirection;thistypeofapproachwasusedunderthecontrolofascriptinglanguagetoaddreal-timegesturalactivitytotheanimatedcharacterOLGA(Beskow,J.andMcGlashan,S.1997).Atthe
otherextreme,theBadlerresearchgroupattheUni-versityofPennsylvaniaarguedthattrulylifelikemo-tionrequirestheuseofalargenumberofparametersconcernedwitheffortandshaping,creatingthean-imationmodelEMOTEbasedonLabanMovementAnalysis;howeverthemodelisnon-emotionalanddoesnotaddressautonomousactiongeneration(Chi,D.,Costa,M.,Zhao,L.andBadler,N.2000).
Oneofthemostwellknownexamplesofmotionparameterizationthathasinspiredextensiveatten-tioninboththeanimationandroboticscommunitiesisthetechniqueof“verbsandadverbs”proposedbyRoseetal.(Rose,C.,Cohen,M.andBodenheimer,B.1998).Inthismethodverbsarespecificbaseac-tionsandadverbsarecollectionsofparametersthatmodifytheverbstoproducefunctionallysimilarmo-toroutputsthatvaryaccordingtothespecificquali-tiestheadverbsweredesignedtoaffect.
Thistechniqueallows,forexample,ananimatortogenerateacontinuousrangeofemotionalexpressionsofaparticularaction,fromsay‘excitedwaving’to‘forlornwaving’,withouthavingtomanuallycreateeveryspecificexampleseparately;instead,asinglebase‘waving’motionwouldbecreated,andthenpa-rameterrangesthatdescribedthevariationfrom‘ex-cited’to‘forlorn’providethemeansofautomaticallysituatinganexamplesomewhereonthatcontinuum.Whiletheessentialapproachisgeneral,itistypi-callyappliedatthelevelofindividualactionsratherthanoverallbehavioraloutputduetothedifficultyofspecifyingasuitableparameterizationofallpossiblemotion.
Similarly,thetechniqueof“morphablemodels”proposedbyGieseandPoggiodescribesmotionex-pressionsintermsofpatternmanifoldsinsidewhichplausible-lookingmotionscanbesynthesizedandde-composedwithlinearparametercoefficients,accord-ingtotheprinciplesoflinearsuperposition(Giese,M.A.andPoggio,T.2000).Intheiroriginalexam-ple,locomotiongaitssuchas‘walking’and‘marching’wereusedasexemplarstodefinetheparameterspace,andfromthistheparameterscouldbere-weightedinordertosynthesizenewgaitssuchas‘limping’.Fur-thermore,anobservedgaitcouldthenbeclassifiedagainstthetrainingexamplesusingleast-squareses-timationinordertoestimateitsrelationshiptotheknownwalkingstyles.
Notsurprisingly,motionparameterizationtech-niquessuchastheabovehavebeenshownsignificantinterestbythesegmentoftheroboticscommunityconcernedwithrobotprogrammingbydemonstra-
tion.Motionparameterizationholdspromiseforthecentralproblemthatthisapproachattemptstosolve:extrapolationfromadiscrete(andideallysmall)setofdemonstratedexamplestoacontinuoustaskcom-petencyenvelope;i.e.,knowingwhattovarytoturnaknownspatio-temporalsequenceintoonethatisfunc-tionallyequivalentbutbetterrepresentscurrentcir-cumstancesthatwerenotprevailingduringtheorig-inaldemonstration.Theroboticsliteratureinthisarea,evenjustconcerninghumanoidrobots,istoobroadtosummarize,butsee(Peters,R.A.II,Camp-bell,C.C.,Bluethmann,W.J.andHuber,E.2003)forarepresentativeexamplethatusesaverbs-adverbsapproachforalearnedgraspingtask,andillustratesthedepthoftheproblem.
Fortunately,theproblemthatbehavioraloverlaysseekstoaddressissomewhatsimpler.Inthefirstplace,thetaskathandisnottotransformaknownactionthatwouldbeunsuccessfulifexecutedunderthecurrentcircumstancesintoonethatnowachievesasuccessfulresult;rather,itistomakemodificationstocertainbodilyposturesduringknownsuccessfulactionstoadegreethatcommunicatesinformationwithoutcausingthosesuccessfulactionstobecomeunsuccessful.Thisdistinctionconferswithittheluxuryofbeingabletochoosemanyparametersinadvanceaccordingtowell-describedhumanposturetaxonomiessuchasthosereferredtoinSection2,al-lowingalgorithmicattentiontothusbeconcentratedontheappropriatequantityandcombinationoftheirapplication.Furthermore,suchasituation,inwhichtheoutcomeevenofdoingnothingatallisatleastthesuccessoftheunderlyingbehavior,hastheaddedadvantagethatunusedbodilyresourcescanbeem-ployedinoverlayservicewithareasonablelevelofassurednessthattheywillnotcausethebehaviortofail.
Secondly,themotionclassificationtask,whereitexistsatall,isnotthecomplexproblemofparam-eterizingmonolithicobservedoutputmotion-posturecombinations,butsimplytoattempttoensurethatsuchconformations,whenclassifiedbythehumanobserver’sbuilt-inparameterizationfunction,willbeclassifiedcorrectly.Inasense,behavioraloverlaysstartwithmotordescriptionsthathavealreadybeenparameterized—intotheinstrumentalbehaviorit-selfandtheoverlayinformation—andthetaskoftheoverlayfunctionistomaintainandapplythisparameterizationinsuchafashionthattheoutputremainseffectiveinsubstanceandnaturalinappear-ance,afarlessambiguoussituationthanthereverse
10
caseofattemptingtoseparateunconstrainednaturalbehaviorintoitsinstrumentalandnon-instrumentalaspects(e.g.‘style’from‘content’).
Inlightoftheabove,andsincebehavioraloverlaysaredesignedwiththeintentionofaffectingalloftherobot’sbehavioralconformations,notjustonesthathavebeendevelopedorexhibitedatthetimeofpa-rameterestimation,theimplementationofbehavioraloverlaysdescribedherehasfocusedontwomainar-eas.First,thedevelopmentofgeneralrulesconcern-ingthedesireddisplaybehaviorsandavailablebodilyresourcesidentifiedinSection2thatcanbeappliedtoawiderangeoftherobot’sactivities.Andsecond,thedevelopmentofamaintenanceandreleasingmecha-nismthatmapstheserulestothespaceofemotionalandotherinternalinformationthattherobotwillusethemtoexpress.Detailsoftheinternaldatarepre-sentationsthatprovidetherobotwiththecontextualinformationnecessarytomaketheseconnectionsaregiveninSection4,andimplementationdetailsoftheoverlaysystemitselfareprovidedinSection5.
Figure5:FunctionalunitsandinterconnectionsinQRIO’sstandardEGOArchitecture.
4RelationshipsandAttitudes
Animportantdrivingforcebehindproxemicsandbodylanguageistheinternalstateoftheindivid-ual.QRIO’sstandardEGOArchitecturecontainsasystemofinternally-maintainedemotionsandstatevariables,andsomeoftheseareapplicabletonon-verbalcommunication.Abriefreviewfollows;pleaserefertoFigure5forablockdiagramoftheunmodi-fiedEGOArchitecture.
QRIO’semotionalmodel(EM)containssixemo-tions(ANGER,DISGUST,FEAR,JOY,SADNESS,SURPRISE)plusNEUTRAL,alongthelinesoftheEkmanproposal(Ekman,P.andDavidson,R.J.1994).CurrentlyQRIOisonlyabletoexperienceoneemotionfromthislistatagiventime,thoughthenon-verbalcommunicationoverlaysystemhasbeendesignedinanticipationofthepotentialforthistochange.Emotionallevelsarerepresentedbycontinuous-valuedvariables.
QRIO’sinternalstatemodel(ISM)isalsoasys-temofcontinuous-valuedvariables,thataremain-tainedwithinacertainrangebyahomeostasismech-anism.ExamplevariablesincludeFATIGUE,IN-FORMATION,VITALITYandINTERACTION.AlowlevelofaparticularstatevariablecanbeusedtodriveQRIOtoseekobjectsoractivitiesthatcanbeexpectedtoincreasethelevel,andviceversa.
FormoredetailsoftheQRIOemotionally
groundedarchitecturebeyondtheabovesummary,see(Sawada,T.,Takagi,T.,Hoshino,Y.andFu-jita,M.2004).Theremainderofthissectiondetailsadditionstothearchitecturethathavebeendevel-opedspecificallytosupportnon-verbalcommunica-tionoverlays.Inthepre-existingEGOarchitecture,QRIOhasbeendesignedtobehavedifferentlywithdifferentindividualhumansbychangingitsEMandISMvaluesinresponsetofacialidentificationofeachhuman.However,thesechangesareinstantaneousuponrecognitionofthehuman;thelackofaddi-tionalfactorsdistinguishingQRIO’sfeelingsaboutthesespecificindividualhumanslimitstheamountofvariationandnaturalnessthatcanbeexpressedintheoutputbehaviors.
InorderforQRIOtorespondtoindividualhu-manswithmeaningfulproxemicandbodylanguagedisplaysduringpersonalinteractions,QRIOrequiresamechanismforpreservingthedifferencesbetweentheseindividualpartners—whatwemightgenerallyrefertoasarelationship.QRIOdoeshavealong-termmemory(LTM)feature;anassociativememory,itisusedtorememberconnectionsbetweenpeopleandobjectsinpredefinedcontexts,suchasthenameofaperson’sfavoritefood.Tosupportemotionalrelationships,whichcanthenbeusedtoinfluencenon-verbalcommunicationdisplay,adatastructuretoextendthissystemhasbeendeveloped.
EachhumanwithwhomQRIOisfamiliarisrepre-sentedbyasingle‘Relationship’structure,withsev-eralinternalvariables.AdiagramofthestructurecanbeseeninFigure6.Thesetofvariableschosenforthisstructurehavegenerallybeenselectedforthepracticalpurposeofsupportingnon-verbalcommuni-cationratherthantomirroraparticulartheoreticalmodelofinterpersonalrelationships,withexceptions
11
Figure6:AnexampleinstantiationofaRelationship
Inasimilarvein,individualhumanscanexhibitstructure,showingarbitrarilyselectedvaluesforthe
singlediscreteenumeratedvariableandeachofthemarkedlydifferentoutputbehaviorundercircum-stancesinwhichparticularaspectsoftheirinternalfivecontinuousvariables.
statescouldbesaidtobeessentiallyequivalent;theirpersonalitiesaffectthewayinwhichtheiremotionsanddesiresareinterpretedandexpressed.Thereisnotedbelow.
Thediscrete-valued‘Type’fieldrepresentsthegen-evidencetosuggestthattheremaybepositiveout-eralnatureofQRIO’srelationshipwiththisindivid-comestoendowinghumanoidrobotswithpercepti-ual;itprovidestheglobalcontextwithinwhichthebleindividualdifferencesinthewayinwhichtheyothervariablesarelocallyinterpreted.Becauseapri-reacttotheirinternalsignalsandtheirrelationshipsmaryuseofthisstructurewillbethemanagementwithpeople.Studiesinsocialpsychologyandcom-ofpersonalspace,ithasbeenbasedondelineationsmunicationhaverepeatedlyshownthatpeoplepre-thatmatchtheprincipalproxemiczonessetoutbyfertointeractwithpeoplesharingsimilarattitudesWeitz(Weitz,S.1974).AnINTIMATErelationshipandpersonalities(e.g.(Blankenship,V.,Hnat,S.M.,signifiesaparticularlyclosefriendlyorfamilialbondHess,T.G.andBrown,D.R.2004,Byrne,D.andinwhichtouchisaccepted.APERSONALrelation-Griffit,W.1969))—suchbehaviorisknownas“sim-shiprepresentsmostfriendlyrelationships.ASO-ilarityattraction”.Acontrarycaseismadefor“com-CIALrelationshipincludesmostacquaintances,suchplementaryattraction”,inwhichpeopleseekinter-astherelationshipbetweenfellowcompanyemploy-actionswithotherpeoplehavingdifferentbutcom-ees.AndaPUBLICrelationshipisoneinwhichplementaryattitudesthathavetheeffectofbalanc-QRIOmaybefamiliarwiththeidentityofthehumaningtheirownpersonalities(Kiesler,D.J.1983,Orford,butlittleornosocialcontacthasoccurred.Itisen-J.1986).visagedthatthisvaluewillnotchangefrequently,butitcouldbelearnedoradaptedovertime(forexam-ple,aPUBLICrelationshipbecomingSOCIALafterrepeatedsocialcontact).
AllotherRelationshipvariablesarecontinuous-valuedquantities,boundedandnormalized,withsomecapableofnegativevaluesifareactionsimi-larinintensitybutoppositeinnatureissemanticallymeaningful.The‘Closeness’fieldrepresentsemo-tionalclosenessandcouldalsobethoughtofasfa-miliarityoreventrust.The‘Attraction’fieldrepre-sentsQRIO’sdesireforemotionalclosenesswiththehuman.The‘Attachment’field,basedonBowlby’stheoryofattachmentbehavior(Bowlby,J.1969),isavariablewithdirectproxemicconsequencesandrep-resentswhetherornotthehumanisanattachmentobjectforQRIO,andifsotowhatdegree.The‘Sta-tus’fieldrepresentstherelativesenseofsuperiorityorinferiorityQRIOenjoysintherelationship,allowingthestructuretorepresentformallyhierarchicalrela-Thesetheorieshavebeencarriedoverintothesphereofinteractionsinvolvingnon-humanpartici-pants.Intheproductdesignliterature,Jordandis-cussesthe“pleasurability”ofproducts;twoofthecategoriesinwhichproductshavethepotentialtosatisfytheirusersare“socio-pleasure”,relatingtointer-personalrelationships,and“ideo-pleasure”,re-latingtosharedvalues(Jordan,P.W.2000).Andanumberofhuman-computerinteractionstudieshavedemonstratedthathumansrespondtocomputersassocialagentswithpersonalities,withsimilarityat-tractionbeingthenorm(e.g.(Nass,C.andLee,K.M.2001)).Yanetal.performedexperimentsinwhichAIBOroboticdogswereprogrammedtosim-ulatefixedtraitsofintroversionorextroversion,andshowedthatsubjectswereabletocorrectlyrecognizetheexpressedtrait;however,inthiscasetheprefer-encesobservedindicatedcomplementaryattraction,withtheauthorspostulatingtheembodimentoftherobotitselfasapotentialfactorinthereversal(Yan,
tionshipsinadditiontoinformalfriendshipsandac-quaintances.Finally,theRelationshipstructurehasa‘Confidence’fieldwhichrepresentsQRIO’sassess-mentofhowaccuratelytheothercontinuousvariablesinthestructuremightrepresenttheactualrelation-ship;thisprovidesamechanismforallowingQRIO’sreactionstoapersontoexhibitdifferingamountsofvariabilityastheirrelationshipprogresses,perhapstendingtosettleasQRIOgetstoknowthembetter.
12
C.,Peng,W.,Lee,K.M.andJin,S.2004).
Asaresult,ifahumanoidrobotcanbestbethoughtasaproductwhichseekstoattractandul-timatelyfulfilthedesiresofhumanusers,itmightbereasonabletopredictthathumanswillbemoreattractedtoandsatisfiedbyrobotswhichappeartomatchtheirownpersonalitiesandsocialresponses.Ontheotherhand,ifahumanoidrobotisinsteadbestdescribedasahuman-likeembodiedagentthatisalreadyperceivedascomplementarytohumansasaresultofitsdifferencesinembodiment,itmightalternativelybepredictedthathumanswilltendtobemoreattractedtorobotshavingpersonalitiesthatareperceptiblydifferentfromtheirown.Ineithercase,arobotpossessingthemeanstoholdsuchat-titudeswouldbelikelytohaveageneraladvantageinattainingacceptancefromhumans,andultimatelythechoiceoftheprecisenatureofanindividualrobotcouldbeleftuptoitshumancounterpart.
ToallowQRIOtoexhibitindividualistic(andthushopefullymoreinteresting)non-verbalbehavior,asystemthatinterpretsorfilterscertainaspectsofQRIO’sinternalmodelwasrequired.Atpresentthissystemisintendedonlytorelatedirectlytotherobot’sproxemicandbodylanguageoverlays;noat-tempthasbeenmadetogivetherobotanythingap-proachingacomplete‘personality’.Assuch,thedatastructurerepresentingtheseindividualtraitsisin-steadentitledan‘Attitude’.EachQRIOhasonesuchstructure;itisenvisagedtohavealong-to-medium-termeffect,inthatitisreasonableforthestructuretobepre-setandnottochangethereafter,butitalsomightbeconsidereddesirabletohavetherobotbeabletochangeitsnaturesomewhatoverthecourseofaparticularlylongterminteraction(evenalifetime),muchashumanattitudessometimesmelloworbe-comemoreextremeovertime.BriefinstantiationoftemporaryreplacementAttitude(andRelationship)structureswouldalsoprovideapotentialmechanismforQRIOtoexpanditsentertainmentrepertoirewith‘acting’ability,simplybyusingitsnormalmecha-nismstorespondtointeractionsasthoughtheyfea-tureddifferentparticipants.
TheAttitudestructureconsistsofsixcontinuous-valued,normalizedvariablesinthreeopposingpairs,asillustratedinFigure7.Opposingpairsaredirectlyrelatedinthatonequantitycanbecomputedfromtheother;althoughonlythreevariablesarethereforecomputationallynecessary,theyarespecifiedinthiswaytobemoreintuitivelygroundedfromthepointofviewoftheprogrammerorbehaviordesigner.
Figure7:AnarbitraryinstantiationofanAtti-tudestructure,showingthethreeopposingpairsofcontinuous-valuedvariables,andtheISM,EMandRelationshipvariablesthateachinterprets.
The‘Extroversion’and‘Introversion’fieldsareadapteddirectlyfromtheExtroversiondimensionoftheFive-FactorModel(FFM)ofpersonality(McCrae,R.R.andCosta,P.T.1996),andasdetailedabovewereusedsuccessfullyinexperimentswithAIBO.InthecaseofQRIO,thesevaluesareusedtoaf-fecttheexpressionofbodylanguageandtointerpretinternalstatedesires.HighvaluesofExtroversionencouragemoreovertbodylanguage,whereashighIntroversionresultsinmoresubtlebodilyexpression.ExtroversionincreasestheeffectoftheISMvariableINTERACTIONanddecreasestheeffectofINFOR-MATION,whereasIntroversiondoesthereverse.The‘Aggressiveness’and‘Timidity’fieldsaremorelooselyadaptedfromtheFFM—theycanbethoughtofassomewhatsimilartoahybridofAgreeable-nessandConscientiousness,thoughtheexactna-tureismorecloselytailoredtothespecificemotionsandnon-verbaldisplayrequirementsofQRIO.Ag-gressivenessincreasestheeffectoftheEMvariableANGERanddecreasestheeffectofFEAR,whileTimidityaccentuatesFEARandattenuatesANGER.HighTimiditymakessubmissiveposturesmoreprob-able,whilehighAggressivenessraisesthelikelihoodofdominantposturesandmaynegatetheeffectofRelationshipStatus.
Finally,the‘Attachment’and‘Independence’fieldsdepartfromtheFFMandreturntoBowlby;theironlyeffectisproxemic,asaninterpreterforthevalueofRelationshipAttachment.Whileanyhumancanrepresentanattachmentrelationshipwiththerobot,robotswithdifferentattitudesshouldbeexpectedto
13
respondtosuchrelationshipsindifferentways.Re-lationshipAttachmentisintendedtohavetheeffectofcompressingthedistancefromthehumanthattherobotiswillingtostray;therobot’sownIndepen-denceorAttachmentcouldbeusedforexampletoalterthefall-offprobabilitiesattheextremesofthisrange,ortochangethe‘sortie’behavioroftherobottobiasittomakebrieferforaysawayfromitsattach-mentobject.
ThescalarvalueswithintheRelationshipandAt-titudestructuresprovidetherawmaterialforaffect-ingtherobot’snon-verbalcommunication(andpo-tentiallymanyotherbehaviors).Thesehavebeencarefullyselectedinordertobeabletodrivetheout-putoverlaysinwhichweareinterested.However,therearemanypossiblewaysinwhichthismaterialcouldthenbeinterpretedandmathematicallycon-vertedintotheoverlaysignalsthemselves.Wedonotwishtoargueforoneparticularnumericalalgorithmoveranother,becausethatwouldamounttoclaimingthatwehavequantitativeanswerstoquestionssuchas“howoftenshouldarobotwhichis90%introvertedand65%timidleanawayfromapersontowhomitisonly15%attracted?”.Wedonotmakesuchclaims.Instead,wewillillustratetheinterpretationofthesestructuresthroughtwoexamplesofgeneraldataus-agemodelsthatcontrastthevariabilityofoverlaygenerationavailabletoastandardQRIOversusthatavailabletooneequippedwithRelationshipsandAt-titudes.
Sociofugal/sociopetalaxis:Weuseascalarvalueforthesociofugal/sociopetalaxis,S,from0.0(thetorsofacingstraightahead)to1.0(maximumoff-axistorsorotation).SincethisrepresentsQRIO’sunwill-ingnesstointeract,aQRIOequippedwithnon-verbalcommunicationskillsbutneitherRelationshipsnorAttitudesmightgeneratethisaxisbasedona(pos-siblynon-linear)functionsofitsISMvariableIN-TERACTION,IINT,anditsEMvariableANGER,EANG:
S=s(IINT,EANG)
TheQRIOequippedwithRelationshipsandAtti-tudesisabletoapplymoredatatowardsthiscompu-tation.Thevalueoftherobot’sIntroversion,AInt,candecreasetheeffectofINTERACTION,resultingininhibitionofdisplayofthesociofugalaxisaccord-ingtosomecombiningfunctionf.Conversely,thevalueoftherobot’sAggression,AAgg,canincreasetheeffectofANGER,enhancingthesociofugalresultaccordingtothecombiningfunctiong.Furthermore,
therobotmaychoosetotakeintoaccounttherelativeStatus,RSta,ofthepersonwithwhomitisinteract-ing,inordertopolitelysuppressanegativedisplay.Theimprovedsociofugalaxisgenerationfunctions+isthusgivenby:
S=s+(f(IINT,AInt),g(EANG,AAgg),RSta)Proxemicdistance:EvenifthesetofappropriateproxemiczonesPforthetypeofinteractionisspeci-fiedbytheinteractionbehavioritself,QRIOwillneedtodecideonanactualscalardistancewithinthosezones,D,tostandfromthehuman.AstandardQRIOmightthususeitsinstantaneousEMvariableFEAR,EFEA,toinfluencewhetheritwaspreparedtochooseaclosevalueortoinsteadactmorewarilyandstandback,accordingtoanotherpossiblynon-linearfunctiond:
D=d(P,EFEA)
TheenhancedQRIO,ontheotherhand,isabletomakemuchmorenuancedselectionsofproxemicdis-tance.TheproxemicTypefieldoftheRelationship,RTyp,allowstherobottoselectthemostappropriatesinglezonefromtheoptionsgiveninP.Thevalueoftherobot’sTimidity,ATim,increasestheeffectofFEAR,alteringtheextenttowhichtherobotdisplayswarinessaccordingtoacombiningfunctionu.TheClosenessoftherobot’srelationshiptothehuman,RClo,canalsobecommunicatedbyalteringthedis-tanceitkeepsbetweenthem,ascanitsAttractionforthehuman,RAttr.IftherobothasanAttachmentre-lationshipwiththehuman,itsstrengthRAttacanbeexpressedbyenforcinganupperboundonD,modu-latedbytherobot’sIndependenceAIndaccordingtothecombiningfunctionv.Theimprovedidealprox-emicdistancegenerationfunctiond+isthusgivenby:
P,RTyp,u(EFEA,ATim),RClo,RAttr,v(RAtta,AInd)
D=d+
Clearly,theadditionofRelationshipsandAtti-tudesofferincreasedscopeforvariabilityofnon-verbalcommunicationsoutput.Moreimportantly,however,theyprovidearich,sociallygroundedframeworkforthatvariability,allowingstraightfor-wardimplementationstobedevelopedthatnon-verballycommunicatetheinformationinawaythatvariespredictablywiththebroadersocialcontextoftheinteraction.
14
5
SystemArchitectureandIm-plementation
TheQRIObehavioralsystem,termedtheEGOAr-chitecture,isadistributedobject-basedsoftwarear-chitecturebasedontheOPEN-Rmodularenviron-mentoriginallydevelopedforAIBO.ObjectsinEGOareingeneralassociatedwithparticularfunctions.Therearetwobasicmemoryobjects:Short-TermMemory(STM)whichprocessesperceptualinforma-tionandmakesitavailableinprocessedformatarateof2Hz,andLong-TermMemory(LTM)whichassociatesinformation(suchasfacerecognitionre-sults)withknownindividualhumans.TheInternalModel(IM)objectmanagesthevariablesoftheISMandEM.TheMotionController(MC)receivesandexecutesmotioncommands,returningaresulttotherequestingmodule.
QRIO’sactualbehaviorsareexecutedinuptothreeSituatedBehaviorLayer(SBL)objects:theNormalSBL(N-SBL)managesbehaviorsthatex-ecuteatthe2HzSTMupdaterate(homeostaticbehaviors);theReflexiveSBL(R-SBL)managesbe-haviorsthatrequireresponsesfasterthantheN-SBLcanprovide,andthereforeoperatesatasignificantlyhigherupdatefrequency(behaviorsrequiringpercep-tualinformationmustcommunicatewithperceptualmodulesdirectlyratherthanSTM);anddeliberativebehaviorcanberealizedintheDeliberativeSBL(D-SBL).
WithineachSBL,behaviorsareorganizedinatree-structurednetworkofschemas;schemasperformminimalcommunicationbetweenoneanother,andcompeteforactivationaccordingtoawinner-take-allmechanismbasedontheresourcerequirementsofin-dividualschemas.FormoreinformationabouttheEGOArchitectureandtheSBLsystemofbehaviorcontrol,pleasesee(Fujita,M.,Kuroki,Y.,Ishida,T.andDoi,T.2003).
Figure8:InterdependencediagramoftheEGOAr-chitecturewithNVC.
Non-VerbalCommunication(NVC)objectwasaddedtotheEGOArchitecture.
TheNVCobjecthasdataconnectionswithseveraloftheotherEGOobjects,butitsmostcentralfunc-tionasamotor-leveloverlayobjectistointerceptandmodifymotorcommandsastheyaresenttotheMCobject.AdditionoftheNVCobjectthereforeinvolvesreconnectingtheMCoutputofthevariousSBLobjectstotheNVC,andthenconnectingtheNVCobjecttotheMC.MCresponsesarelikewiseroutedthroughtheNVCobjectandthenbacktotheSBLs.
Inadditiontothemotorconnection,theNVCob-jectmaintainsaconnectiontotheIMoutput(forreceivingISMandEMupdates,whichcanalsobeusedasa2Hzinterrupttimer),theSTMtargetup-date(foracquiringinformationaboutthelocationofhumans,usedinproxemiccomputations)andacus-tommessagechanneltotheN-SBLandR-SBL(forreceivingspecialinformationabouttheinteraction,andsendingtriggermessagesforparticulargesturesandpostures).SeeFigure8foragraphicaloverviewoftheEGOArchitecturewithNVC.Inadditiontotheconnectionsshown,theNVCobjectmanagesthebehavioraloverlayvaluesthemselveswithreferencetotheRelationshipandAttitudestructures;Figure9hasanoverviewoftheinternalworkingsofNVC.
5.1NVCObject
Becausebehavioraloverlaysmustbeabletobeap-pliedtoallbehaviorschemas,andbecauseschemasareintendedtoperformminimaldatasharing(sothatschematreescanbeeasilyconstructedfromin-dividualschemaswithouthavingtobeawareoftheoveralltreestructure),itisnotpossibleordesirabletocompletelyimplementanappropriateoverlaysys-temwithintheSBLsthemselves.Instead,toimple-mentbehavioraloverlaysanadditional,independent
15
rameterizedmotioncommandhavingthesameactu-atorresourcerequirements(thispossibilityisnotyettakenadvantageofinthecurrentimplementation).ResourcemanagementintheEGOArchitectureiscoarsegrainedandfixed;itisusedformanagingschemaactivationintheSBLsaswellasjustforpre-sentingdirectmotioncommandconflicts.There-sourcecategoriesinEGOaretheHead,Trunk,RightArm,LeftArmandLegs.Thusabodylanguagemotioncommandthatwantedonlytoadjusttheso-ciofugal/sociopetalaxis(trunkrotate),forexample,wouldneverthelessbeforcedtotakecontroloftheentiretrunk,potentiallyblockinganinstrumentalbe-haviorfromexecuting.TheNVCobjectimplementsasomewhatfiner-grainedresourcemanagerbyvirtueoftheoverlaysystem.Bybeingabletochoosetomodifythecommandsofinstrumentalbehaviorsdi-rectly,ornottodoso,theeffectisasiftheresourceweremanagedatthelevelofindividualjoints.Thisraises,however,theproblemofthecaseoftimestepsatwhichenvironmentalbehaviorsdonotsendmotioncommandsyetitisdesiredtomodifytherobot’sbodylanguage;thissituationiscommon.TheNVCobjectskirtsthisproblembyrelyingonanetworkoffast-activatingidleschemasresidingintheN-SBL.EachidleschemaisresponsibleforasingleMCresource.Oneachtimestep,ifnoinstrumen-talbehaviorclaimsagivenresource,theappropriateidleschemasendsanullcommandtotheNVCob-jectforpotentialoverlaying.Ifanoverlayisdesired,thenullcommandismodifiedintoagenuineactionandpassedontotheMC;otherwiseitisdiscarded.SincecommandsfromidleandotherspecialoverlayschemasarethustreateddifferentlybytheNVCob-jectthanthosefrombehavioralschemas,theNVCobjectmustkeeptrackofwhichschemasarewhich;thisisaccomplishedbyahandshakingprocedurethatoccursatstartuptime,illustratedgraphicallyinFig-ure10.Normaloperationoftheidleschemasisillus-tratedinFigure11.
ForpracticalconsiderationswithintheEGOAr-chitecture,actionsinvolvedinnon-verbalcommuni-cationcanbedividedintofourcategoriesaccordingtotwoclassificationaxes.Firstisthetimingrequire-mentsoftheaction.Someactions,suchasposturalshifts,arenotpreciselytimedandareappropriatelysuitedtothe2HzupdaterateoftheN-SBL.Oth-ers,however,arehighlycontingentwiththeactivity,suchasnoddingduringdialog,andmustresideintheR-SBL.Secondistheresourcerequirementsoftheaction.Someactions,suchasgrossposture,re-
Figure9:Thenon-verbalcommunication(NVC)module’sconceptualinternalstructureanddatain-terconnectionswithotherEGOmodules.
5.2Overlays,ResourcesandTiming
Thedatarepresentationforthebehavioraloverlaysthemselvesisbasicyetflexible.TheyaredividedaccordingtothemajorresourcetypesHead,Arms,TrunkandLegs.Foreachjoint(orwalkingparametervalue,inthecaseofthelegs)withinaresourcetype,theoverlaymaintainsavalueandaflagthatallowsthevaluetobeinterpretedaseitheranabsolutebias(toallowconstantchangestotheposturalconforma-tion)orarelativegain(toaccentuateorattenuateincomingmotions).Inaddition,eachoverlaycate-gorycontainsatimeparameterforalteringthespeedofmotionoftheresource,whichcanalsobeflaggedasabiasoragain.Finally,thelegsoverlaycontainsanadditionalegocentricpositionparameterthatcanbeusedtomodifythedestinationoftherobotinthecaseofwalkingcommands.
MotioncommandsthatareroutedthroughtheNVCobjectconsistofuptotwoparts:acommandbody,andanoptionparameterset.Motionsthatareparameterized(i.e.,thathaveanoptionpart)canbemodifieddirectlybytheNVCobjectaccordingtothecurrentvaluesoftheoverlaysthattheNVCob-jectisstoring.Suchtypesofmotionsincludedirectpositioningofthehead,trunkandarmwithexplicitjointanglecommands;generalpurposemotionsthathavebeendesignedwithreuseinmind,suchasnod-ding(theparameterspecifyingthedepthofnod);andcommandsthatareintendedtosubsequentlyrunindirectcommunicationwithperceptualsystemswiththeSBLexcludedfromthedecisionloop,suchasheadtracking.UnfortunatelyduetothedesignoftheMCsystem,unparameterizedmotioncommands(i.e.,thosewithjustabody)cannotbealteredbeforereachingtheMCobject;buttheycanbeignoredorreplacedwithanyothersingleparameterizedorunpa-
16
Figure11:Normalidleactivityofthesystemintheabsenceofgestureorreflextriggering.N-SBLidleandgestureschemasareactiveifresourcespermit.ActiveidleschemassendnullmotorcommandstoNVConeachtimestep.Thedialogreflexparentisal-waysactive,buttheindividualdialogreflexschemasareinactive.
Figure10:SpecialschemasregisterwiththeNVCobjectinahandshakingprocedure.Anacknowledge-mentfromNVCisrequiredbecausetheSBLsdonotknowwhentheNVCobjectisreadytoacceptmessages,sotheyattempttoregisteruntilsuccess-ful.WithintheR-SBL,theparentregistersinsteadoftheindividualschemas,topreserveascloselyaspossiblethemodeofoperationofthepre-existingconversationalreflexsystem.Theregistrationsys-temalsosupportsallowingschemastodeliberatelychangetypeatanytime(e.g.fromanidleschematoabehavioralschema)thoughnouseofthisextensi-bilityhasbeenmadetodate.
quireonlypartialcontrolofaresource,whereasoth-ersrequireitstotalcontrol.Partialresourcemanage-menthasjustbeendescribedabove,anditfunctionsidenticallywithcommandsthatarealsosynchronousandoriginateintheR-SBL.Howevertherearealsonon-verbalcommunicationbehaviorsthatrequireto-talresourcecontrol,suchasemblematicgestures,andthesearealsoimplementedwiththeuseofspecial-ized“triggered”schemas:attheN-SBLleveltheseschemasarecalledgestureschemas,andattheR-SBLleveltheyarecalleddialogreflexschemas.
GestureschemasresidewithinthesamesubtreeoftheN-SBLastheidleschemas;unliketheidleschemas,however,theydonotattempttoremainactivewhennoinstrumentalbehaviorisoperating.Instead,theyawaitgesturetriggermessagesfromtheNVCobject,becausetheNVCobjectcannotcreatemotioncommandsdirectly,itcanonlymodifythem.Uponreceivingsuchamessage,agestureschemade-
17
Figure13:TriggeringofadialogreflexoccursbyNVCsendinganactivationmessagetotheactivedialogreflexparent,whichthensignalstheself-evaluationfunctionsofitschildren.Childrenself-evaluatingtoactive(e.g.speechisoccurring)raisetheirownactivationlevelsandsendmotorcommandsifrequired.Thedialogreflexsystemremainsactiveandperiodicallyself-evaluatinguntilreceivingadeac-tivationmessagefromNVC,somayeitherbetightlysynchronizedtonon-causaldata(e.g.pre-marked
terminesifitisdesignedtoexecutetherequestedoutgoingspeech)orgenerallyreactivetocausaldatagesture;ifso,itincreasesitsactivationlevel(AL),(e.g.incomingspeech).sendsthemotioncommand,andthenreducesitsAL
again(Figure12).Thegesture,whichmaybeanun-values,ratherthanthetriggeringobjectdoingitforparameterizedmotionfromapredesignedset,isthenthem.InadditiontotheexistingheadnoddingdialogexecutedbytheMCobjectasusual.reflex,aseconddialogreflexschemawasaddedtobeDialogreflexschemasoperateinasimilarbutresponsibleforleaningtherobot’storsointowardsslightlydifferentfashion.ExistingresearchonQRIOthespeaker(ifthestateoftherobot’sinterestandhasalreadyresultedinasuccessfulreactivespeechtheRelationshipdeemitappropriate)inreactiontointeractionsystemthatinsertedcontingentattentivespeech(Figure13).headnodsandspeechfilleractionsintotheflowof
conversation(Aoyama,K.andShimomura,H.2005).5.3ExecutionFlowItwasofcoursedesirablefortheNVCsystemtocom-plement,ratherthancompetewith,thisexistingsys-TheoverallflowofexecutionoftheNVCbehavioraltem.Thepriorsystemusedthe“intention”mech-overlaysystem(Figure14)isthusasfollows:anismtomodifytheALsofdialogreflexschemas,
•Atsystemstartup,theidle,andgestureschemas
residingintheR-SBL,atappropriatepointsinthe
mustregisterthemselveswiththeNVCobject
dialog.TheNVCobjectmanagesthesereflexesin
toallowsubresourceallocationandcommunica-almostexactlythesameway.
tion.TheseschemassettheirownALhightoal-Dialogreflexschemasaregroupedunderaparentlowthemtosendmessages;whentheyreceiveanschemawhichhasnoresourcerequirementsandisal-acknowledgementfromNVC,theyreducetheir
waysactive.UponreceivingacontrolmessagefromALstoalevelthatwillensurethatinstrumentaltheNVCobject,theparentschemaactivatesitschil-behaviorstakepriority(Figure10).
dren’smonitorfunctions,whichthenchoosetoin-creasetheirownALs,triggerthereflexgestureand•Alsoatsystemstartup,thedialogreflexschemathenreducetheirALsagain.Thustheonlypracti-parentregistersitselfwithNVCsothatNVC
caldifferencefromthepreviousmechanismisthatknowswheretosenddialogreflextriggermes-thereflexschemasthemselvesmanagetheirownALsages(Figure10).Figure12:TriggeringofagestureoccursbyNVCsendingatriggermessagetoallregisteredN-SBLschemas.Thereisnodistinctionbetweenidleandgestureschemasasidleschemascanincludegesture-likeresponsessuchasposturalshifts.Individualac-tiveschemasdecodethetriggermessageandchoosewhetherornottorespondwithamotorcommand;inthistypicalexampleonegestureschemaresponds.
18
mandsmayholdaresourceuntilrequestedbyanbehaviorcommand,butcompletingbehaviorcommandsfreetheresourceimmediately.•Tomaketherobot’sappearancemorelifelikeandlessrigid,theregularISMupdateisusedtotriggerperiodicre-evaluationoftheupperbodyoverlays.Whenthisoccurs,resourcesheldbyidlecommandsarefreedandthenextincomingidlecommandsareallowedtoexecutethenewoverlays.•STMupdatesaremonitoredforchangesintheproxemicstatecausedbymovementofthetargethuman,andifnecessarytheoverlaysareregen-eratedtoreflecttheupdatedsituation.Thispro-videsanopportunityfortriggeringoftherobot’semblematicgesturesconcernedwithmanipulat-ingthehumanintochangingtheproxemicsit-uation,ratherthantherobotadjustingitdi-rectlywithitsownlocomotion.Aprobabilisticfunctionisexecutedthatdependsmainlyontherobot’sIMandAttitudeandthepriorresponsetosuchgestures,andaccordingtotheresultsanemblematicgestureoftheappropriatenatureandurgency(e.g.beckoningthehumancloser,orwavinghimorheraway)istriggered.Trig-germessagesaresenttoallregisteredidleandgestureschemas,withituptotheschemastochoosewhetherornottoattempttoexecutetheproposedaction(Figure12).•ActivationmessagesaresentfromNVCtothedialogreflexschemaswhendictatedbythepres-enceofdialogmarkersorthedesireforgeneralreactivityoftherobot(Figure13).
Figure14:ExecutionflowoftheNVCoverlaysystem.Timingandsynchronizationissuessuchasrespond-ingtochangesintheinternalmodelandproxemicstate(fromSTM)andwaitingforthehumantore-spondtoanemblematicgestureareimplicitlyman-agedwithintheoverlaycalculationstate(i.e.enter-ingthisstatedoesnotguaranteethatanychangetotheoverlayswillinfactbemade).
•Proxemicandbodylanguageactivitycom-menceswhenanN-SBLschemainformsNVCthatanappropriateinteractionistakingplace.PriortothiseventNVCpassesallinstrumentalmotioncommandstoMCunalteredanddiscardsallidlecommands(Figure11).TheN-SBLinter-actionbehavioralsopassestheIDofthehumansubjectoftheinteractiontoNVCsoitcanlookuptheappropriateRelationship.
•NVCgeneratesoverlayvaluesusingasetoffunc-tionsthattakeintoaccounttheIMandAttitudeoftherobot,theRelationshipwiththehumanandtheproxemichintoftheinteraction(ifavail-able).Thisoverlaygenerationphaseincludestherobot’sselectionoftheidealproxemicdis-tance,whichisconvertedintoawalkingoverlayifneeded.
•Idlecommandsbegintobeprocessedaccordingtotheoverlayvalues.Asidlecommandstyp-icallyneedonlybeexecutedonceuntilinter-rupted,theNVCobjectmaintainsatwo-tieredinternalsubresourcemanagerinwhichidlecom-
6Results
Allofthenon-verbalcommunicationcapabilitiessetoutinSection2wereimplementedaspartoftheover-laysystem.Totesttheresultingexpressiverangeoftherobot,weequippedQRIOwithasuiteofsam-pleAttitudesandRelationships.TheAttitudeselec-tionscomprisedatimid,introvertedQRIO,anag-gressive,agonisticQRIO,andan“average”QRIO.TheexampleRelationshipswerecraftedtorepresentan“oldfriend”(dominantattribute:highCloseness),an“enemy”(dominantattribute:lowAttraction),andthe“SonyPresident”(dominantattribute:highStatus).Thetestscenarioconsistedofasimpleso-cialinteractioninwhichQRIOnoticesthepresence
19
ofahumanintheroomandattemptstoengagehimorherinaconversation.Theinteractionschemawasthesameacrossconditions,buttheactiveAtti-tudeandRelationship,aswellasQRIO’scurrentEMandISMstate,wereconcurrentlycommunicatednon-verbally.Thisscenariowasdemonstratedtolabora-torymembers,whowereabletorecognizetheappro-priatechangesinQRIO’sbehavior.DuetothehighdegreeofexistingfamiliarityoflaboratorymemberswithQRIO,however,wedidnotattempttocollectquantitativedatafromtheseinteractions.
DuetotheconstraintsofourarrangementwithSonytocollaborateontheirQRIOarchitecture,for-maluserstudiesofthepsychologicaleffectivenessofQRIO’snon-verbalexpressions(asdistinctfromthedesignofthebehavioraloverlaymodelitself)havebeenrecommendedbutnotyetperformed.Inlieuofsuchexperiments,theresultsofthisworkareil-lustratedwithacomparisonbetweengeneralinter-actionswithQRIO(suchastheexampleabove)inthepresenceandabsenceoftheoverlaysystem.LetusspecifythebaselineQRIOasequippedwithitsstandardinternalsystemofemotionsandinstincts,aswellasbehavioralschematreesfortrackingofthehuman’shead,foraugmentingdialogwithillustratorssuchasheadnoddingandtorsoleaning,andfortheconversationitself.ThebehavioraloverlayQRIOisequippedwiththesesameattributesplustheoverlayimplementationdescribedhere.
WhenthebaselineQRIOspotsthehuman,itmaybegintheconversationactivitysolongasithassuf-ficientinternaldesireforinteraction.Ifso,QRIOcommencestrackingthehuman’sfaceandrespond-ingtothehuman’sheadmovementswithmovementsofitsownheadtomaintaineyecontact,usingQRIO’sbuilt-infacetrackingandfixationmodule.WhenQRIO’sbuilt-inauditoryanalysissystemdetectsthatthehumanisspeaking,QRIOparticipateswithitsreflexiveillustratorssuchasheadnodding,sothehu-mancanfindQRIOtoberesponsive.However,itisuptothehumantoselectanappropriateinter-personaldistance—ifthehumanwalkstotheothersideoftheroom,orthrustshisorherfaceupclosetoQRIO’s,theconversationcontinuesasbefore.Sim-ilarly,thehumanhasnoevidenceofQRIO’sdesireforinteractionotherthantheknowledgethatitwassufficientlyhightoallowtheconversationalbehav-iortobeinitiated;andnoinformationconcerningQRIO’sknowledgeoftheirrelationshipotherthanwhatmightcomeupintheconversation.
Contrastthisscenariowiththatofthebehavioral
Figure15:AselectionofQRIObodyposturesgener-atedbytheoverlaysystem.Fromlefttoright,toptobottom:normal(sociopetal)standingposture;maxi-mumsociofugalaxis;defensivearmcrossingposture;submissiveattentivestandingposture;open,outgo-ingraisedarmposture;defensive,pugnaciousraisedarmposture.
overlayQRIO.WhenthisQRIOspotsthehumanandtheconversationactivitycommences,QRIOimmedi-atelybeginstocommunicateinformationaboutitsinternalstateandtherelationshipitshareswiththehuman.QRIObeginstoadoptcharacteristicbodyposturesreflectingaspectssuchasitsdesireforin-teractionanditsrelativestatuscomparedwiththatofthehuman—forsomeexamplesseeFigure15.Ifthehumanistoocloseortoofaraway,QRIOmaywalktoamoreappropriatedistance(seeFigure3forspecificsconcerningtheappropriateinteractiondis-tancesselectedforQRIO).Alternatively,QRIOmaybeckonthehumancloserormotionhimorheraway,andthengivethehumanachancetorespondbeforeadjustingthedistanceitselfifheorshedoesnot.Thismayoccuratanytimeduringtheinteraction—approachtooclose,forexample,andQRIOwillbackawayorgesticulateitsdispleasure.RepeatedcasesofQRIO’sgesturesbeingignoredreducesitspatienceforrelyingonthehuman’sresponse.
ThebehavioraloverlayQRIOmayalsorespondtothehuman’sspeechwithparticipatoryillustrators.Howeveritsbehaviorisagainmorecommunicative:ifthisQRIOdesiresinteractionandhasapositiverela-tionshipwiththehuman,itsnodsandtorsoleanswillbemorepronounced;conversely,whendisengagedoruninterestedinthehuman,theseillustratorswillbeattenuatedorsuppressedentirely.Byadjustingthe
20
parametersofitsheadtrackingbehavior,theoverlaysystemallowsQRIOtoobservablyalteritsrespon-sivenesstoeyecontact.Thespeedofitsmovementsisalsoaltered:anangryorfearfulQRIOmovesmorerapidly,whileaQRIOwhoseinterestintheinterac-tioniswaninggoesthroughthemotionsmoreslug-gishly.AllofthisdisplayedinternalinformationismodulatedbyQRIO’sindividualattitude,suchasitsextroversionorintroversion.Thehuman,alreadytrainedtointerpretsuchsignals,isabletocontinu-ouslyupdatehisorhermentalmodelofQRIOandtherelationshipbetweenthem.
Itisclearthatthelatterinteractiondescribedaboveisricherandexhibitsmorevariation.Ofcourse,manyoftheapparentdifferencescouldequallywellhavebeenimplementedthroughspecifi-callydesignedsupportwithintheconversationalin-teractionschematree.Howevertheoverlaysystemallowsthesedistinguishingfeaturestobemadeavail-abletoexistingbehaviorssuchastheconversationschema,aswellasfutureinteractions,withoutsuchexplicitdesignworkonarepeatedbasis.
Forthemostpart,theoverlaysgeneratedwithinthisimplementationwereabletomakeuseofcontinuous-valuedmappingfunctionsfromthevar-iousinternalvariablestotheposturalandproxemicoutputs.Forexample,acontinuouselementsuchasthesociofugal/sociopetalaxisiseasilymappedtoaderivedmeasureofinteractiondesire(seeSection4).However,therewereanumberofinstancesinwhichitwasnecessarytoselectbetweendiscreteoutputstates.Thiswasduetothehigh-levelmotioncontrolinterfaceprovidedtous,whichsimplifiedthebasicprocessesofgeneratingmotionsonQRIOandaddedanadditionallayerofphysicalsafetyfortherobot,suchaspreventionoffallingdown.Ontheotherhand,thismadesomecapabilitiesofQRIO’sunder-lyingmotionarchitectureunavailabletous.Someexpressiveposturessuchasdefensivearmcrossing,forinstance,didnotlendthemselvestocontinuousmodulationasitcouldnotbeguaranteedviatheab-stractionofthehigh-levelinterfacethattherobot’slimbswouldnotinterferewithoneanother.Inthesecasesadiscretedecisionfunctionwithaprobabilisticcomponentwastypicallyused,basedonthevariablesusedtoderivethecontinuousposturaloverlaysandwithitsoutputallowedtoreplacethoseoverlayvaluesdirectly.
toire,asthehighlevelinterfacewasnotdesignedforparameterizationofatomicgestures.(Thiswaspresumablyasafetyfeature,ascertainparameteriza-tionsmightaffecttherobot’sstabilityandcauseittofallover,evenwhentherobotisphysicallycapableofexecutingthegestureoverthemajorityofthepa-rameterspace.)Thegesturescreatedwereemblem-aticgesturesinsupportofproxemicactivity,suchasanumberofbeckoninggesturesusingvariouscombi-nationsofoneorbotharmsandthefingers.SimilarprobabilisticdiscretedecisionfunctionswereusedtodeterminewhenthesegesturesshouldbeselectedtooverridecontinuouspositionoverlaysforQRIOtoad-justtheproxemicstateitself.
QRIO’shigh-levelwalkinginterfaceexhibitedasimilartrade-offinallowingeasygenerationofbasicwalkingmotionsbutcorrespondinglyhidingaccesstounderlyingmotionparametersthatcouldotherwisehavebeenusedtoproducemoreexpressiveposturesandcomplexlocomotionbehaviors.BodylanguageactionsthatQRIOisotherwisecapableofperform-ing,suchasstandingwithlegsakimbo,ormakingfinely-tunedproxemicadjustmentsbywalkingalongspecificcurvedpaths,werethusnotavailabletousatthistime.Wethereforeimplementedanalgorithmthatusedtheavailablelocomotionfeaturestogen-erateproxemicadjustmentsthatappearedasnatu-ralaspossible,suchasuninterruptiblelinearwalkingmovementstoappropriatepositions.Thedisplaybe-haviorsimplementedinthisprojectshouldthusbeviewedasarepresentativesampleofamuchlargerbehaviorspacethatcouldbeoverlaidonQRIO’smo-toroutputatlowerlevelsofmotioncontrolabstrac-tion.
7
ConclusionsWork
andFuture
Thisresearchpresentedthemodelofabehavioraloverlaysystemanddemonstratedthatanimplemen-tationofthemodelcouldsuccessfullybeusedtomodulatetheexpressivenessofbehaviorsdesignedwithoutnon-verbalcommunicationinmindaswellasthosespecificallycreatedtonon-verballysupportverbalcommunication(e.g.dialogillustrators).De-spitelimitationsinthemotioncontroller’ssupportformotionparameterization,aspectrumofdisplay
Inadditiontodiscretearmpostures,itwassimi-behaviorswasexhibited.FutureextensionstothelarlynecessarytocreateseveralexplicitgestureswithmotioncontrollercouldsupportincreasedflexibilityamotioneditorinordertoaugmentQRIO’sreper-inthecommunicativeabilitiesofthesystemwithout
21
majoralterationstothebasicoverlayimplementa-Finally,returningtofirstprinciples,thispaper
tion.soughttoarguethatsinceitiswellsupportedthatTheproxemiccomponentofthebehavioraloverlaynon-verbalcommunicationisanimportantmodeofsystememphasizedthatthespatialnatureofnon-humaninteraction,itwillalsobeausefulcomponentverbalcommunicationthroughthemanagementofofinteractionbetweenhumansandhumanoidrobots,interpersonaldistanceisnotonlyanimportantfea-andthatbehavioraloverlayswillbeanadequateandtureofhuman-robotinteractionthatislargelyyetscalablemethodoffacilitatingthisinthelongterm.tobeexplored,butcanalsobeeffectivelymodulatedRobustHRIstudiesarenownecessarytoconfirmorbymeansofanoverlay.Theproxemicsinthiscaserejectthisconjecture,sothatfuturedesignworkcanwerecomputedsimplyintermsofegocentriclinearproceedinthemosteffectivedirections.distances,sothereremainsplentyofpotentialforexploringalternativeconceptualizationsofproxemicoverlays,suchasforcefieldsordeformablesurfaces.Inordertobetrulyproxemicallycommunicative,fu-tureworkisrequiredingivingQRIOabetterspatialunderstandingofitsenvironment,notonlyincludingphysicalfeaturessuchaswallsandobjects,butalsoamoredetailedmodelofthespatialactivitiesofthehumanswithinit.
8Acknowledgements
TheworkdescribedinthisdocumentwasperformedatSonyIntelligenceDynamicsLaboratories(SIDL)duringasummerresearchinternshipprogram.TheauthorswishtogratefullyacknowledgethegeneroussupportandassistanceprovidedbyYukikoHoshino,KuniakiNoda,KazumiAoyama,HidekiShimomura,
TheoverlaygeneratingfunctionsusedinthistherestofthestaffandmanagementofSIDL,andtheprojectwereeffectivedisplaygeneratorsbyvirtueothersummerinternshipstudents,intherealizationofbeinghand-tunedtoreflectthegeneralbodyofofthisresearch.researchintohumanpostures,gesturesandinterper-sonalspacing,withaccommodationsmadeforthedif-ferencesinbodysizeandmotioncapabilityofQRIO.ReferencesAssuch,theyrepresentasystemthatisgenerically
applicablebutinflexibletoculturalvariationsandAlthaus,P.,Ishiguro,H.,Kanda,T.,Miyashita,
T.andChristensen,H.I.:2004,Navigationforthelike.Theoverlaygenerationcomponentsthus
human-robotinteractiontasks,Proc.Interna-offeranopportunityforfutureeffortstoincorporate
tionalConferenceonRoboticsandAutomation,learningandadaptationmechanismsintotherobot’s
Vol.2,pp.14–1900.communicativerepertoire.Inparticular,itwouldbe
beneficialtobeabletoextractmorereal-timefeed-Aoyama,K.andShimomura,H.:2005,Realworldbackfromthehuman,especiallyinthecaseofthespeechinteractionwithahumanoidrobotonproxemics.Forexample,giventhenatureofcurrentalayeredrobotbehaviorcontrolarchitecture,human-robotinteraction,beingabletorobustlyde-Proc.InternationalConferenceonRoboticsand
tectahuman’sdiscomfortwiththeproxemicsoftheAutomation.situationshouldbeanimportantpriority.
Thisworkintroducedsignificantadditionalinfras-Bailenson,J.N.,Blascovich,J.,Beall,A.C.and
Loomis,J.M.:2003,InterpersonaldistanceintructureconcerningQRIO’sreactionstoitsinter-immersivevirtualenvironments,Personalityandnalstatesandemotionsandtospecifichumans,in
SocialPsychologyBulletin29,819–833.theformoftheAttitudeandRelationshipstructures.
TheRelationshipsupportassistsinprovidinganemo-Beskow,J.andMcGlashan,S.:1997,OLGA—a
tionalandexpressivecomponenttoQRIO’smemory,conversationalagentwithgestures,Proc.IJCAI-andtheAttitudesupportoffersthebeginningofan97WorkshoponAnimatedInterfaceAgents:individualisticaspecttotherobot.ThereisagreatMakingThemIntelligent.dealofpotentialforfutureworkinthisarea,primar-ilyinthemanagementandmaintenanceofrelation-Bethel,C.L.andMurphy,R.R.:2006,Affectiveshipsonceinstantiated;andin“closingtheloop”byexpressioninappearance-constrainedrobots,allowingthehuman’sactivity,asviewedbytherobotProceedingsofthe2006ACMConferenceonthroughthefilterofthespecificrelationship,tohaveHuman-RobotInteraction(HRI2006),SaltLakeafeedbackeffectontherobot’semotionalstate.City,Utah,pp.327–328.
22
Blankenship,V.,Hnat,S.M.,Hess,T.G.andBrown,Dittmann,A.:1978,Theroleofbodymove-D.R.:2004,Reciprocalinteractionandsimilar-mentincommunication,NonverbalBehavior
ityofpersonalityattributes,JournalofSocialandCommunication,LawrenceErlbaumAsso-andPersonalRelationships1,415–432.ciates,Hillsdale.Bowlby,J.:1969,Attachment&LossVolume1:At-Ekman,P.andDavidson,R.J.:1994,TheNatureof
Emotion,OxfordUniversityPress.tachment,BasicBooks.Fong,T.,Nourbakhsh,I.andDautenhahn,K.:2003,
Breazeal,C.:2000,SociableMachines:ExpressiveAsurveyofsociallyinteractiverobots,Robotics
SocialExchangeBetweenHumansandRobots,andAutonomousSystems42,143–166.PhDthesis,MassachusettsInstituteofTechnol-Fridlund,A.:1994,HumanFacialExpression:Anogy.
EvolutionaryView,AcademicPress,SanDiego,
Breazeal,C.:2003,Towardssociablerobots,RoboticsCA.
andAutonomousSystems42(3–4),167–175.
Fujita,M.,Kuroki,Y.,Ishida,T.andDoi,T.:2003,
Autonomousbehaviorcontrolarchitectureofen-Breazeal,C.,Brooks,A.G.,Gray,J.,Hoffman,G.,
tertainmenthumanoidrobotSDR-4X,Proc.In-Kidd,C.,Lee,H.,Lieberman,J.,Lockerd,A.
ternationalConferenceonIntelligentRobotsandandChilongo,D.:2004,Tutelageandcollabora-Systems,LasVegas,NV,pp.960–967.tionforhumanoidrobots,InternationalJournalGiese,M.A.andPoggio,T.:2000,Morphablemodelsfortheanalysisandsynthesisofcomplexmotion
Brooks,A.G.,Berlin,M.,Gray,J.andBreazeal,
pattern,InternationalJournalofComputerVi-C.:2005,Untetheredroboticplayforrepetitive
sion38(1),59–73.
physicaltasks,Proc.ACMInternationalConfer-enceonAdvancesinComputerEntertainment,Gordon,R.:1986,Folkpsychologyassimulation,Valencia,Spain.MindandLanguage1,158–171.Byrne,D.andGriffit,W.:1969,SimilarityandGuye-Vuilleme,A.,Capin,T.K.,Pandzic,I.S.,Thal-mann,N.M.andThalmann,D.:1998,Non-awarenessofsimilarityofpersonalitycharacter-verbalcommunicationinterfaceforcollaborativeisticdeterminantsofattraction,JournalofEx-virtualenvironments,Proc.CollaborativeVir-perimentalResearchinPersonality3,179–186.
tualEnvironments(CVE98),pp.105–112.
Cassell,J.andVilhjalmsson,H.:1999,Fullyembod-Hall,E.T.:1966,TheHiddenDimension,Doubleday,
iedconversationalavatars:Makingcommunica-GardenCity,NY.
tivebehaviorsautonomous,AutonomousAgents
Hara,F.,Akazawa,H.andKobayashi,H.:2001,andMultiagentSystems2,45–.
RealisticfacialexpressionsbySMAdrivenface
Chi,D.,Costa,M.,Zhao,L.andBadler,N.:robot,ProceedingsofIEEEInternationalWork-2000,TheEMOTEmodelforeffortandshape,shoponRobotandHumanInteractiveCommu-Proc.27thAnnualConf.onComputerGraphicsnication,pp.504–510.
andInteractiveTechniques(SIGGRAPH’00),
Heal,J.:2003,UnderstandingOtherMindsfromthepp.173–182.
Inside,CambridgeUniversityPress,Cambridge,UK,pp.28–44.Christensen,H.I.andPacchierotti,E.:2005,Embod-iedsocialinteractionforrobots,inDautenhahn,Jordan,P.W.:2000,DesigningPleasurableProducts:
K.(ed.),Proceedingsofthe2005ConventionofanIntroductiontotheNewHumanFactors,Tay-theSocietyfortheStudyofArtificialIntelligencelor&FrancisBooksLtd.,UK.andSimulationofBehaviour(AISB-05),Hert-Kanda,T.,Hirano,T.,Eaton,D.andIshiguro,H.:fordshire,England,pp.40–45.
2004,Interactiverobotsassocialpartnersand
Davies,M.andStone,T.:1995,MentalSimulation,peertutorsforchildren,Human-ComputerIn-BlackwellPublishers,Oxford,UK.teraction19,61–84.
23
ofHumanoidRobots1(2),315–348.
Kiesler,D.J.:1983,The1982interpersonalcircle:A
taxonomyforcomplementarityinhumantrans-actions,PsychologicalReview90,185–214.
ConferenceonRoboticsandAutomation,Taipei,Taiwan,pp.2806–2811.
Reeves,B.andNass,C.:1996,TheMediaEquation:
Knapp,M.:1972,NonverbalCommunicationinHu-HowPeopleTreatComputers,Television,and
manInteraction,ReinhartandWinston,Inc.,NewMediaLikeRealPeopleandPlaces,Cam-NewYork.bridgeUniversityPress,Cambridge,England.Kopp,S.andWachsmuth,I.:2000,Aknowledge-Rose,C.,Cohen,M.andBodenheimer,B.:1998,basedapproachforlifelikegestureanimation,Verbsandadverbs:Multidimensionalmotionin-Proc.ECAI-2000.terpolation,IEEEComputerGraphicsandAp-plications18(5),32–40.Kopp,S.andWachsmuth,I.:2002,Model-basedani-mationofcoverbalgesture,Proc.ComputerAn-Sawada,T.,Takagi,T.,Hoshino,Y.andFujita,M.:imation.2004,Learningbehaviorselectionthroughinter-actionbasedonemotionallygroundedsymbolLikhachev,M.andArkin,R.C.:2000,Roboticcom-concept,Proc.IEEE-RAS/RSJInt’lConf.onfortzones,ProceedingsofSPIE:SensorFusion
HumanoidRobots(Humanoids’04).andDecentralizedControlinRoboticSystemsIII
Conference,Vol.4196,pp.27–41.
Machotka,P.andSpiegel,J.:1982,TheArticulate
Body,Irvington.
Smith,C.:2005,Behavioradaptationforasocially
interactiverobot,Master’sthesis,KTHRoyalIn-stituteofTechnology,Stockholm,Sweden.
McCrae,R.R.andCosta,P.T.:1996,TowardateBoekhorst,R.,Walters,M.,Koay,K.L.,Dauten-hahn,K.andNehaniv,C.:2005,Astudyofnewgenerationofpersonalitytheories:Theoret-asinglerobotinteractingwithgroupsofchil-icalcontextsforthefive-factormodel,inWig-dreninarotationgamescenario,Proc.6thIEEEgins,J.S.(ed.),Five-FactorModelofPersonal-InternationalSymposiumonComputationalIn-ity,Guilford,NewYork,pp.51–87.
telligenceinRoboticsandAutomation(CIRA
Nakauchi,Y.andSimmons,R.:2000,Asocialrobot2005),Espoo,Finland.
thatstandsinline,Proc.InternationalConfer-enceonIntelligentRobotsandSystems,Vol.1,Thorisson,K.R.:1996,CommunicativeHumanoids:pp.357–3.AComputationalModelofPsychosocialDia-logueSkills,PhDthesis,MITMediaLaboratory,Nass,C.andLee,K.M.:2001,Doescomputer-Cambridge,MA.
generatedspeechmanifestpersonality?experi-mentaltestofrecognition,similarity-attraction,Walters,M.L.,Dautenhahn,K.,Koay,K.L.,Kaouri,andconsistence-attraction.,JournalofExperi-C.,teBoekhorst,R.,Nehaniv,C.L.,Werry,I.
mentalPsychology,Applied7(3),171–181.andLee,D.:2005a,Closeencounters:Spa-tialdistancesbetweenpeopleandarobotOrford,J.:1986,Therulesofinterpersonalcom-ofmechanisticappearance,Proceedingsofthe
plementarity:Doeshostilitybegethostilityand
5thIEEE-RASInternationalConferenceon
dominance,submission?,PsychologicalReview
HumanoidRobots(Humanoids’05),Tsukuba,
93,365–377.
Japan,pp.450–455.
Pacchierotti,E.,Christensen,H.I.andJensfelt,P.:
2005,Human-robotembodiedinteractioninWalters,M.L.,Dautenhahn,K.,teBoekhorst,R.,
Koay,K.L.,Kaouri,C.,Woods,S.,Nehaniv,hallwaysettings:Apilotuserstudy,Proceed-C.L.,Lee,D.andWerry,I.:2005b,Theinfluenceingsofthe2005IEEEInternationalWorkshop
ofsubjects’personalitytraitsonpersonalspa-onRobotsandHumanInteractiveCommunica-tialzonesinahuman-robotinteractionexperi-tion,pp.1–171.
ment,ProceedingsofIEEERo-Man2005,14th
Peters,R.A.II,Campbell,C.C.,Bluethmann,W.J.AnnualWorkshoponRobotandHumanInter-andHuber,E.:2003,Robonauttasklearn-activeCommunication,IEEEPress,Nashville,
ingthroughteleoperation,Proc.InternationalTennessee,pp.347–352.
24
Weitz,S.:1974,NonverbalCommunication:Read-ingsWithCommentary,OxfordUniversityPress.Yamasaki,N.andAnzai,Y.:1996,Activeinter-faceforhuman-robotinteraction,Proc.Interna-tionalConferenceonRoboticsandAutomation,pp.3103–3109.Yan,C.,Peng,W.,Lee,K.M.andJin,S.:2004,Can
robotshavepersonality?Anempiricalstudyofpersonalitymanifestation,socialresponses,andsocialpresenceinhuman-robotinteraction,thAnnualConferenceoftheInternationalCommu-nicationAssociation.Zecca,M.,Roccella,S.,Carrozza,M.C.,Cappiello,
G.,Cabibihan,J.-J.,Dario,P.,Takanobu,H.,Matsumoto,M.,Miwa,H.,Itoh,K.andTakan-ishi,A.:2004,Onthedevelopmentoftheemo-tionexpressionhumanoidrobotWE-4RIIwithRCH-1,Proc.IEEE-RAS/RSJInt’lConf.onHumanoidRobots(Humanoids’04),LosAnge-les,California.
25
因篇幅问题不能全部显示,请点此查看更多更全内容
Copyright © 2019- ovod.cn 版权所有 湘ICP备2023023988号-4
违法及侵权请联系:TEL:199 1889 7713 E-MAIL:2724546146@qq.com
本站由北京市万商天勤律师事务所王兴未律师提供法律服务