Embodied Agents for Multi-party Dialogue in Immersive Virtual Worlds
David Traum
University of Southern California Institute for Creative Technologies
13274 Fiji Way Marina del Rey, CA 90292
[email protected]
Jeff Rickel
University of Southern California Information Sciences Institute
4676 Admiralty Way Marina del Rey, CA 90292
[email protected]
ABSTRACT
Wepresentamo delofdialogue foremb o diedvirtualagents
that can communicate with multiple (human and virtual)
agentsinamulti-mo dal setting, including face-to-facesp o-
ken and nonverbal, as well as radio interaction, spanning
multiple conversations in supp ort of an extended complex
task.Themo delbuildsonpreviousworkinemb o diedagents
andmulti-layerdialoguemo dels, andisb eingdeployed ina
p eacekeepingmissionrehearsal exercisesetting.
1. INTRODUCTION
Immersive virtual worlds oer exciting p otential for rich
interactive exp eriences. Human users can cohabit three-
dimensionalgraphicalenvironmentswithvirtualhumansfor
entertainment, education,andtraining. Theycanhavead-
ventures in fantasy worlds. They canlearn ab out history
orother cultures by exp eriencing life indistant places and
times. They can practice tasks, make mistakes, and gain
exp erience without the consequences of real-world failure.
In all these applications , virtual humans can play a wide
variety of roles, including mentorsand guides, teammates,
companions,adversaries, andthelo cal p opulace.
Perhapsthegreatestchallengeincreatingvirtualhumansfor
interactive exp eriencesis supp orting face-to-facecommuni-
cation among p eople and virtual humans. On one hand,
virtual worlds are an ideal application forcurrent sp oken
languagetechnology: theyprovide amicroworldwherecon-
versation can legitimately b e restricted to the events and
objects within its connes. On the other hand, they raise
issues that have received relatively little attention in com-
putational linguisti cs. First,face-to-face communication in
virtualworldsrequiresattentiontoallthenonverbalsignals
(e.g., gaze, gestures, and facial displays) that accompany
humansp eech.Second,conversationsthataresituatedina
3Dworldraiseahostofissues,includingtheattentional fo-
cusoftheconversants,whetherandtowhatdegreetheycan
seeandhearoneanother,andtherelativelo cationsofcon-
versantsand theobjectsthey arediscussing. Finally, since
therewill typicallyb emultiplerealandvirtualp eople, vir-
tual worlds require supp ort for multi-party conversations,
including theability toreasonab outtheactiveparticipants
in a conversation as well as who else might b e listening.
While there has b een some early work in the area ofem-
b o diedconversationalagents[10,19],andsomeofthiswork
hasaddressedhuman-agentdialogues situatedin3Dvirtual
worlds[31],thereiscurrently nogeneralmo delofsuchdia-
logues.
Inthispap er,wediscussprogresstowardsamo delofmulti-
partydialogueinimmersivevirtualworlds. Ourmo delbuilds
onpriorworkonemb o diedconversationalagents,theso cial
psychology literature onthe nonverbal signals that accom-
pany human sp eech, and mo dels of collab orati ve dialogue
from computational linguistics . Themo del is organized as
asetof dialogue management layers, eachincluding an in-
formation stateandasetofdialogue actsthatchangethat
state. The layers include traditional ones, such as turn-
takingandgrounding,aswellasseveralnovellayersaddress-
ing theissues ofmulti-party dialogue in immersive worlds.
Physicalandverbalactionswilloftencontributetomultiple
layers.
2. EXAMPLE SCENARIO
The test-b ed for our emb o died agents is the Mission Re-
hearsalExerciseprojectatTheUniversityofSouthernCal-
ifornia'sInstituteforCreativeTechnologies. Thesettingfor
this project is a virtual reality theatre, including a visual
sceneprojectedontoan8fo ottallscreenthatwrapsaround
the viewer in a 150 degree arc (12 fo ot radius). Immer-
sive audio software provides multiple tracks of spatialized
sounds,playedthroughtensp eakerslo catedaroundtheuser
andtwosubwo ofers.
Within this setting, a virtual environment has b een con-
structed, representing a small village in Bosnia, complete
withbuilding s,vehicles,andvirtualcharacters. Withinthis
environment has b een placed an Army p eacekeeping sce-
nario: aU.S.Armylieutenantndshimselfinthepassenger
seatofasimulatedArmyvehiclesp eedingtowardsaBosnian
village tohelpaplato on introuble. Suddenly,herounds a
andamedic.
into acivilian vehicle, injuring alo cal b oy (Figure1). The
b oy'smotherandanArmymedicarehunchedoverhim,and
asergeantapproachesthelieutenanttobriefhimonthesit-
uation. Urgentradiocallsfromtheotherplato on,aswellas
o ccasional explosions andweap onsrefromthatdirection,
suggest that the lieutenant send his tro ops to help them.
Emotional pleas from theb oy'smother, as well as a grim
assessmentbythemedicthattheb oyneedsamedevacim-
mediately,suggestthatthelieutenantinsteadusehistro ops
tosecurealandingzoneforthemedevachelicopter.
A preliminary demonstration simulation has b een imple-
mented [36], using Hollywo o d storytelling, combined with
technical exp ertise from USC. Figure 2 shows a small ex-
cerptfromthesimulation script.
Inthis interaction, a numb er of issues arise foremb o died
agents, some going b eyond capabiliti es of current imple-
mentedsystems. First,atabroadlevel,wecanseethatthe
agentsconcernthemselveswithmultiple agentsand multi-
pleconversationsduringthisinteraction. Themainscenein
Figure2concerned theLtandSgt,butthemedic wasalso
brought in, and the mother was an imp ortant overhearer.
Other plato on memb ers and townsp eople may also b e p o-
tential overhearers. There is also a separate conversation
b etweenthePlato onSgtandthesquadleaders,starting at
theend of the excerpt given here. Also, in other partsof
thescenario,theLtengagesinradioconversationswithhis
homebase,anotherplato on,and sometimesamedevache-
licopter. Some oftheseconversations havexed b eginning
andendingp oints(esp eciallytheradioconversations),while
othersaremoreepiso dic,trailingawayasthelo calpurp ose
oftheinteractionisestablished andresolved,andattention
of the parties shifts to other matters. In all cases, agents
mustreasonab outwhotheyaretalkingto,whoislistening,
andwhethertheyareb eingaddressedornot.
Thereisalsoanissue,intheimmersivevirtualworld,ofco-
ordination of sp eechwithother communicative mo dalities.
Inmanycases,gesturesandothernonverbalcuesareimp or-
tantincarryingsomeofthecommunicativefunction. Some
examples herearethewaytheSergeant walksuptotheLt
to initiate conversation, the way thatthe Sergeant glances
atthemedic tosignal thatheshouldtaketheturnandre-
sp ondtotheLt'squestion,andthewaythemedicglancesat
themotherwhileformulating alessdirectanswerab outthe
b oy'shealth|fo cusingontheconsequenceofhiscondition
ratherthandirectlystatingwhatmight b eupsetting tothe
mother.
3. PRIOR WORK
Our work builds on prior work in the areas of emb o died
conversational agents[10]andanimatedp edagogical agents
[19]. Several systems have carefully mo deled theinterplay
b etween sp eech and nonverbal b ehavior [9,11, 8, 28],but
thesesystems havefo cused exclusively on dyadic conversa-
tion, and theydid notallowusers andagents tocohabit a
virtualworld. TheGandalfsystem[11]allowedanagentand
humantocohabitarealphysicalspace,andtousegazeand
gestureto referencean object(i.e.,awall-mounted display
screen) inthat space,butthe agent'spresence was limited
to ahead andhandon a2D computermonitor. Similarl y,
the Reaagent [8]cantransp ort herselfto and into virtual
houses and apartments, and the user can p oint to some
objects within those virtual environments, but the user is
notimmersedintheenvironment,andRea'smovementand
references within those environments is verylimited. The
Cosmo agent [24]includes a sophisticated sp eech and ges-
turegenerationmo dulethatcho osesappropriatedeicticref-
erencesandgesturestoobjectsinitsvirtualworldbasedon
b oth spatial considerations and the dialogue context, but
the agent and itsenvironment are renderedin 2D and the
userdo es notcohabitthevirtualworldwithCosmo.
In contrast, Steve [31, 33, 32] cohabits 3D virtual worlds
with p eople and other Steve agents, so it has addressed
b othmulti-party and immersive asp ects ofdialogue in vir-
SGT WalkuptoLT
LT Sergeant,whathapp enedhere?
SGT Theyjustshotoutfromthesidestreetsir. Gesturingtowardsthecivilia n vehicle
Thedrivercouldn'tsee'emcoming.
LT Howmanyp eoplearehurt?
SGT Theb oy andoneofourdrivers. Gesturingtowardtheb oy
LT Aretheinjuriesserious?
SGT Makeseyecontactwithmedicandno ds
MEDIC Driver'sgotacrackedribbutthe kid's{ Glancing atthemother
Sir,wegottagetaMedevacinhereASAP.
LT We'll getit.
LT Plato on Sergeant,securethearea.
SGT YesSir!
SGT (Shouting) Squadleaders! Raises arm
Listenup! Lo oks aroundatsquadleaders
Figure2: Multi-mo dal,multi-characterinteraction excerpt (many nonverbalb ehaviors omitted).
move around invirtual worlds,they aresensitive towhere
humanusersarelo oking,theycanusegazeanddeicticges-
turestoreferencearbitraryobjectsinthoseworlds,andthey
canusegaze toregulate turn-taking in multi-party (team)
dialogues. However, while Steveincludes adialogue mo del
builtonideasfromcomputationallinguisti cs[33],itfallsfar
shortofthe mo delsin state-of-the-artsp okendialogue sys-
tems. Moreover,themo delfo cusesprimarilyonthecontext
ofdyadic conversations b etween aSteve agent andhis hu-
manstudent;thereisverylittledialoguecontextmaintained
forthemulti-partydialogues b etweenaSteveagentandhis
humanandagentteammates.
Workincomputationallinguisti cshasfo cusedonacomple-
mentarysetofissues: ithaslargelyignoredissuesofemb o di-
mentandimmersioninvirtualworlds,buthaspro ducedrel-
ativelysophisticatedmo delsofsp okendialoguethatinclude
a variety of ho oks formultiple mo dalities. We follow the
frameworkof theTrindi project [23],using dialoguemoves
(in this case, corresp onding to actions) as abstract input
andoutput descriptionsforthedialogue mo delling comp o-
nent. This serves particularly well for considering multi-
mo dalcommunication,sinceitallowsmaximumexibilityof
description,includin gmovesthatcouldb eambiguouslyreal-
izedusingeithersp eechoranothermo dality,movesthatre-
quirerealizationusingacombinationofmultiplemo dalities,
ormovesthatsp ecifyamo dality. Wealsoviewthedialogue
moves (and theaÆliated information states)as segmented
into anumb er of layers, eachconcerning a distinct asp ect
ofinformation state,and usingdierent classesof dialogue
acts. Moreover, thereisno oneto onecorresp ondence b e-
tweendialogueactsandatomiccommunicationrealizations:
a single utterance (or gestural communicative action) will
generally corresp ond to multiple (parts of) dialogue acts,
and it may take several communications (sometimes split
into multiple mo dalities) torealizesomedialogue acts.
Asastartingp oint,weusethedialoguelayersdevelop ed in
theTRAINSandEDISdialoguesystems[39,29,26].These
includedlayersforturn-taking,grounding,coresp eechacts,
andargumentationacts(latertermedforwardandbackward-
lo okingacts[16]). Whilenotfullyimplemented withinnat-
ural language dialogue systems, therehas also b een some
otherworkonotherlayersthatb ecomeimp ortant fordeal-
ing with the multi-character, multi-conversati on domain.
ThisincludesworkbyNovickonmeta-lo cutionary acts,in-
cluding anattention level[27],workby Allwo o dandClark
on basiccommunicative functions [2,13],workbyBunton
interaction management functions [7], and work on multi-
levelgroundinginanextendedmulti-mo daltaskinteraction
[15].
4. CURRENT WORK:
MULTI-MODAL DIALOGUE MODEL
Our dialogue agents are designed to run within the Mis-
sionRehearsal Exerciseenvironment [36,30].Thisenviron-
mentincludesamessage-passingeventsimulator,immersive
sound,andgraphicsincludin gstaticsceneelementsandsp e-
cialeects,renderedbyMultigen/Paradi gm's Vega.
Ouragentmo delisbasedonSteve,asdescrib edin thepre-
vious section. Within the Mission Rehearsal Exercise sce-
nario, Steve (and other agents) are given dynamically an-
imated b o dies fromBoston Dynamics'PeopleShop [6];the
primitive motions werecreatedusing motion capture, and
theSteveagentssequencethesemotions dynamicallyin re-
sp onse to the situation by sending commands to the Peo-
pleShoprun-timesoftware. Themedicandsergeantinclude
expressivefacescreatedby Haptek(www.haptek.com)that
supp ortsynchronizatio noflipmovementstosp eech.Steve's
dialoguemo delandrepresentationoftheinteractional state
isb eing augmentedwiththenewdialoguemo del presented
here.
Our dialogue mo del currently consists of the layers shown
in Figure3. Eachoftheseismo deledfromthep ersp ective
ofanagent involved intheinteraction.
We will rst briey describ e each of these, and then give
detailsofthelayersandhowtheasso ciated actsmayb ere-
alizedusingthepaletteofmulti-mo dalcommunicativeabili-
ties. Thecontactlayerconcernswhetherandhowotherindi-
vidualscanb eaccessibleforcommunication. Mo dalities in-
attention
conversation
{ participants
{ turn
{ initiative
{ grounding
{ topic
{ rhetorical
so cialcommitments(obligation s)
negotiation
Figure3: Multi-party, Multi-conversationDialogue
Layers
attentionlayer concerns the object or pro cess that agents
attendto. Contactisaprerequisite forattention. TheCon-
versationlayermo dels the separatedialogue episo des that
goonthroughouttheinteraction. Aconversationisareied
pro cessentity,consisting ofanumb erofsub-elds. Eachof
theseeldsmayb edierentfordierentconversationshap-
p ening at the same time. The participantsmay b e active
sp eakers,addressees,oroverhearers[14].Theturnindicates
theparticipantwiththerighttocommunicate(usingthepri-
marychannel). Theinitiativeindicatestheparticipant who
is managing the content expressed. The groundingcom-
p onent of a conversation tracks how information is added
to thecommon ground ofthe participants. Theconversa-
tion structure alsoincludes a topicthatgoverns relevance,
andrhetoricalconnectionsb etweenindividu alcontentunits.
Once material is grounded, even as it still relates to the
topic and rhetorical structure of an ongoing conversation,
itisalso addedtothe so cialfabriclinking agents,andnot
partofanyconversation. Thisincludessocialcommitments
|b othobligation s toact orrestrictions on action, aswell
ascommitmentstofactualinformation. Thereisalsoane-
gotiationlayer,mo dellinghowagentscometoagreeonthese
commitments. Wenowturntothelayersinmoredetail.
Thecontactlayerismo delled asavectorforallparticipants
that the agent may interact with, eachelement indicating
whethertheparticipant isincontactin themediasp ecied
ab ove. Therearealsodimensionsforwhethersomeoneisin
contacttosendorreceivecommunications bythismo dality.
Theactionsinuencing thislayeraremake-contact,which
couldb e established by turningon aradio orwalkingover
towithineyecontactorearshot,andbreak-contact,which
couldb eestablished bywalkingoutofhearing,turningout
of view (or moving b ehind something), or turning o the
radio. Contact is not generally realized verbally, although
onemightindicateadesireforcontact,e.g.,byshoutingfor
someonetocomeover. Anexampleofamake-contactaction
isshownattheb eginningofourexampleinFigure2,where
theSergeantwalksovertothelieutenant,toinitiatecontact
ofcontact,thoughalsoincludin g an entry fortheagent it-
self, and attention is a one-way phenomenon, rather than
having (p otentially ) distinct send and receive dimensions.
Theactions aecting this layerare divided into those that
an agent p erformsconcerning its ownattention, andthose
relatedtotheattentionofotheragents. Give-attentionin-
volves paying attentionto somepro cess,p erson, or object,
aswellassignallin gthisattention. Thiscanb eaccomplished
b othverbally (e.g.,saying\yes")ornonverbally (gazing at
the object of attention). Withdraw-attention removes
the current object from the attention entry of the agent.
It canb eimplicit ingiving attention to something else, or
p erformed explicitly, bylo oking away insome cases(other
than whenengagedinconversation andserving someother
purp ose,suchasplanningaturnorindicatingturn-taking).
Request-attention signals to an agent that its attention
is desired { it will also require a give-attention action by
the other agent to change the attentional state. Request-
attentioncanb esignalled byacallontheradio,ashout,or
usinganagent'sname,butalsobygestures,suchasraising
an armorwaving. Arelease-attentionactindicates that
attention isno longer required. It o ccurs by default when
apro cess or action thatis the objectofattention ends. It
canalso b e explicit, in the form of adismissal, or gesture
indicatinglackofattention(lo okingaway).Attentionofthe
released agent maystill p ersist, however, until withdrawn,
orgiventosomething else. Direct-attentionsignals that
attentionshouldb egiventoanotherobjectorevent,rather
than the signaller. This canoften b eaccomplished witha
deicticgesture,orwithanutterancesuchas\lo okup!"
Conversation is often a purp ose forwhich attention is de-
sired. Inthiscase,attentionwillb eassumed(unlessexplic-
itlywithdrawn)forthedurationoftheconversation. There
are also explicit indicators of conversational op enings and
closings [34,22]. Conversations are oftenop ened withver-
bal greetings,butnonverbal actionscanb e veryimp ortant
as well. Kendon found a variety of nonverbal actions in-
volvedintheinitiationofconversation[22].Theinteraction
typicallystartswitha\sighting"b eforetheorientationand
approach. Individual s who do not know each other well
and have no sp ecial reason togreet eachother will \catch
the eye" by gazing longer than normal. Next, a \distance
salutation" involves denite eye contact. This is followed
by an approach, which typically involves gaze avoidance.
Finally, a \close salutation" involves resumed eye contact.
TheBo dyChatsystem[12]wasthersttomo deltheseacts
in animated agents. Conversational op enings and closings
areveryformalizedinthemilitaryradiomo dality,e.g.,say-
ing \out" to close a conversation. Actions suchas op en,
continueandcloseaconversationcanb ep erformedeither
explicitly orimplicitly. Also,thereareactionsformaintain-
ing and changing the presence and status of participants.
Anexamplefromthesampledialogue isthewaythemedic
isbroughtintotheconversationfromb einganoverhearerto
anactiveparticipant.
Turn-taking actionsmo del shiftsin the turnholder. Most
canb erealizedverbally,nonverbally,orthroughacombina-
tion ofthetwo. Take-turnisan attempttotaketheturn
bystartingtosp eak. Request-turn(e.g.,signalled byvar-
ious sp eechpreparation signals suchasop ening the mouth
sp eaker's gaze at phrase b oundaries) is an attempt to re-
questtheturnwithoutforciblytaking it[3].Release-turn
(e.g., signalled by an intonational b oundary tone,removal
ofthe handsfromgesturespace,orasustained gazeatthe
listener at the end ofan utterance) isan attempt to oer
theturnto the listener [3,17]. Hold-turn(e.g.,signalled
verballybyalledpause,ornonverballybygazeaversionat
phraseb oundariesorhandsingesturespace)isanattempt
tokeeptheturnatap oint where thelistenermight other-
wisetake it[3,17]. Thesefour turn-taking actshaveb een
mo deledin emb o died conversational agentssincethe earli-
estsystems [9]. Inmulti-party dialogue, thereisonemore
act: assign-turn (e.g.,signalled verbally byavo cativeex-
pressionornonverballybyasp eaker'sgazeattheendofan
utterance)canb eusedtoexplicitly selectthenextsp eaker
[3]. Among emb o died conversationalagents,only Stevein-
cludesthisact.
WeuseInitiativetomo del howtheagent shouldplancon-
tributions. Even though the turn may shift from sp eaker
to sp eaker, in many parts of a dialogue, a single agent is
controlling the owof the contributions, while others only
resp ond to the initiati ve of the other. For some mixed-
initiativedialogues, initiative may shift from one partici-
pant to another. Initiative is sometimes pre-allo cated by
role forsp ecic tasks. Otherwise, it startswith the agent
who op ened the conversation, and can b e managed with
take-initiative, hold-initiative, and release-initiative
actions. These acts can often b e signalled by p erforming
(only)appropriate core-sp eechactsincontext,e.g.,as pro-
p osedby[41,40].Wearenotcurrentlyconsidering nonver-
balsignalsofinitiative.
Weusethegroundingactsfrom[39,37]:initiate;continue;
repair;request-repair;display;acknowledge;request-
acknowledge,andcancel.Seepreviousworkfordetailsof
allbutdisplay,whichisanexplicitsignalofwhatwasunder-
sto o d(e.g., rep eating aword,p erformingan action), leav-
ingitto theoriginal sp eakerto decideifthis actfunctions
asarepair, request-repair,or acknowledge [20]. Emb o died
conversationalagentstypicallyincludenonverbalactionsfor
request-acknowledge (e.g., gaze at listener at grammatical
pauses) and acknowledge (e.g., gaze at sp eaker and no d),
andsomeincluderequest-repair (e.g.,whensp eechrecogni-
tionfails,Peedytheparrotcupshiswingtohisearandsays
\Huh?" [5]).
Topic actions include start-topicand end-topic. Topic
structurecanalsob ecomplex,whennewtopicsarestarted
b eforeoldoneshavecompleted. Topicshiftsofvarioussorts
can b e signalled explicitly with cue phrases (e.g., \now,"
\anyway"),butalsowithnonverbalcues. Headmovements
cansignaltopicshifts;arelatedsequenceofutterancesbya
sp eakertypically uses thesame basicheadmovement, and
thesp eakerwilloftenemployanewtyp eofheadmovement
tomarkthestart ofanewtopic[21]. Topic shiftsarealso
frequentlyaccompaniedbyshiftsinthesp eaker'sb o dyp os-
ture[21].
Our mo del also includes layers for rhetorical structure of
topicalelementswithinaconversation [25],obligations and
commitments[38,1],andnegotiation[4,35]. Wewill,how-
b ecausetheyhavereceivedafairamountofattentioninthe
previousliterature.
5. PLANS AND EVALUATION
Asidefromanalysis ofthescriptedinteraction, wealsohave
twovenuesforevaluatingthemulti-mo dal,multi-party,multi-
conversation dialoguemo del presented ab ove. First,weare
currentlydevelopingaprototyp eend-to-endsp oken-language
dialogue agent (including alsothe emotionmo dels of[18]),
to b e deployed within the immersive virtual world. The
goal for the prototyp e is to handle at least the fragment
presented in Figure 2, and suitable variations. Given ap-
propriate knowledge, goals, andconnections to thevirtual
b o dy, this agent system should b e able to function as ei-
therthesergeant,medic,ormother(weassume,atleastfor
thepresent,ahumanlieutenant, thesubjectofthetraining
exercise).
We are alsogathering actual p erformance datato comple-
menttheHollywo o d-scriptedinteraction. Wehavedesigned
avariant ofthe mission rehearsal exercisewithacombina-
tionofhumanparticipantsandvirtualcharacters|humans
p ortray not just the lieutenant (who is an external test-
subject),butalsotheSergeant,andotheragentswho com-
municate byradio (commandp ost,other plato on,medevac
pilot). Moreover, thereis a\Wizard" controlling the com-
putational agents(and communicating withthe liveones),
usingaxedpaletteofb ehaviors. Initialpilotshavedemon-
strated the viability of the approach, and wewill so on b e
gatheringusabledataonhowthehumanparticipantsinter-
actwitheachotherandwiththevirtualcharacters. Wewill
use this data to evaluate the dialogue mo dels and inform
futuresystemdevelopment.
Acknowledgements
We would like to thank memb ers of ICT and ISIworking
on the MRE project for interesting discussions ab out the
complexities in this scenario. The work describ ed in this
pap er was supp orted in part by the U.S. Army Research
OÆce under contract #DAAD19-99-C-0046. The content
ofthisarticledo esnotnecessarily reectthep ositionorthe
p olicy of theUS Government, and no oÆcialendorsement
shouldb einferred.
6. REFERENCES
[1] J.Allwo o d.Obligations andoptionsin dialogue.
ThinkQuarterly,3:9{18,1994.
[2] J.Allwo o d,J.Nivre,andE.Ahlsen.Onthesemantics
andpragmaticsoflinguisti cfeedback.Journalof
Semantics,9,1992.
[3] M.ArgyleandM.Co ok.GazeandMutualGaze.
Cambridge UniversityPress,Cambridge,1976.
[4] M.Baker.Amo delfornegotiationin
teaching-learni ng dialogues. JournalofArticial
IntelligenceinEducation,5(2):199{254, 1994.
[5] G.Ball,D. Ling,D.Kurlander,J.Miller,D.Pugh,
T.Skelly,A.Stankosky, D.Thiel,M.vanDantzich,
Agents.AAAI/MITPress,MenloPark,CA,1997.
[6] BostonDynamics.PeopleShop1.4UserManual,2000.
[7] H.Bunt.Interactionmanagement functions and
contextrepresentationrequirements.InProceedingsof
theTwenteWorkshoponLanguageTechnology:
DialogueManagementinNaturalLanguageSystems
(TWLT11),pages187{198,1996.
[8] J.Cassell,T.Bickmore,L.Campb ell,
H.Vilhjalmsson, andH.Yan.Conversationasa
systemframework: Designingemb o diedconversational
agents.InJ.Cassell,J.Sullivan ,S.Prevost,and
E.Churchill,editors, EmbodiedConversational
Agents.MITPress,Cambridge,MA,2000.
[9] J.Cassell,C.Pelachaud, N.Badler,M.Steedman,
B.Achorn,T.Becket, B.Douville,S.Prevost,and
M.Stone.Animated conversation: Rule-based
generationoffacial expression,gestureandsp oken
intonationformultiple conversationalagents.In
ProceedingsofACMSIGGRAPH'94,pages413{420,
Reading,MA,1994.Addison-Wesley.
[10] J.Cassell,J.Sullivan, S.Prevost,andE.Churchill,
editors.EmbodiedConversationalAgents.MITPress,
Cambridge,MA,2000.
[11] J.CassellandK.R.Thorisson.Thep owerofano d
andaglance: Envelop evs.emotional feedback in
animated conversationalagents.AppliedArticial
Intelligence,13:519{538,1999.
[12] J.CassellandH.Vilhjalmsson.Fully emb o died
conversationalavatars: Makingcommunicative
b ehaviors autonomous.AutonomousAgentsand
Multi-AgentSystems,2:45{64,1999.
[13] H.H.Clark.Managing problemsinsp eaking. Speech
Communication,15:243{250,1994.
[14] H.H.Clark.UsingLanguage.Cambridge University
Press,Cambridge, England,1996.
[15] P.Dillenb ou rg,D. Traum,andD.Schneider.
Groundinginmulti-mo daltask-oriented collab oration .
InProceedingsoftheEuropeanConferenceonAIin
Education,1996.
[16] DiscourseResourceInitiative.Standards fordialogue
co dinginnaturallanguage pro cessing.Rep ortno.167,
Dagstuhl-Seminar, 1997.
[17] S.Duncan,Jr.Somesignals andrulesfortaking
sp eaking turnsinconversations.InS.Weitz,editor,
NonverbalCommunication,pages298{311.Oxford
UniversityPress,1974.
[18] J.GratchandS.Marsella. Tearsandfears: Mo deling
emotionsandemotional b ehaviors insyntheticagents.
InProceedingsoftheFifthInternationalConference
Animated p edagogical agents:Face-to-faceinteraction
ininteractive learningenvironments.International
JournalofArticialIntelligenceinEducation,
11:47{78,2000.
[20] Y.Katagiri andA.Shimojima. Display actsin
grounding negotiations.InProceedingsofGotalog
2000,the4thWorkshopontheSemanticsand
PragmaticsofDialogue,pages195{198,2000.
[21] A.Kendon.Some relationshipsb etweenb o dymotion
andsp eech.InA.SiegmanandB.Pop e,editors,
StudiesinDyadicCommunication,pages177{210.
PergamonPress,NewYork,1972.
[22] A.Kendon.Adescription ofsomehuman greetings.In
R.MichaelandJ.Cro ok,editors, Comparative
EcologyandBehaviorofPrimates,pages591{668.
Academic Press,NewYork,1973.
[23] S.Larsson andD.Traum.Informationstateand
dialogue managementintheTRINDIdialoguemove
engineto olkit. NaturalLanguageEngineering,
6:323{340,Septemb er2000.Sp ecial IssueonSp oken
Language DialogueSystemEngineering.
[24] J.C.Lester,J.L.Vo erman,S.G.Towns,andC.B.
Callaway.Deictic b elievabil ity: Co ordinating gesture,
lo comotion, andsp eechinlifelikep edagogical agents.
AppliedArticialIntelligence,13:383{414,1999.
[25] W.C.MannandS.A.Thompson.Rhetorical
structure theory: Atheoryoftextorganization.
Technical Rep ortISI/RS-87-190,USC,Information
SciencesInstitute, June1987.
[26] C.Matheson,M.Po esio,andD.Traum.Mo delling
grounding anddiscourseobligation s usingup date
rules.InProceedingsoftheFirstConferenceofthe
NorthAmericanChapteroftheAssociationfor
ComputationalLinguistics,2000.
[27] D.Novick.ControlofMixed-InitiativeDiscourse
ThroughMeta-LocutionaryActs: AComputational
Model.PhDthesis,University ofOregon,1988.also
available asU.OregonComputer andInformation
Science TechRep ortCIS-TR-88-18.
[28] C.Pelachaud, N.I.Badler,andM.Steedman.
Generating facial expressionsforsp eech.Cognitive
Science,20(1),1996.
[29] M.Po esioandD.R.Traum.Towardsan
axiomatizatio n ofdialogue acts.InProceedingsof
Twendial'98,13thTwenteWorkshoponLanguage
Technology:FormalSemanticsandPragmaticsof
Dialogue,1998.
[30] J.Rickel,J.Gratch,R.Hill,S.Marsella, and
W.Swartout.Stevego estoBosnia: Towardsanew
generation ofvirtualhumansforinteractive
exp eriences.InAAAISpringSymposiumonArticial
IntelligenceandInteractiveEntertainment,March
pro cedural traininginvirtualreality: Perception,
cognition, andmotorcontrol. AppliedArticial
Intelligence,13:343{382,1999.
[32] J.Rickel andW.L.Johnson.Virtualhumansforteam
traininginvirtualreality.InProceedingsoftheNinth
InternationalConferenceonArticialIntelligencein
Education,pages578{585.IOSPress,1999.
[33] J.RickelandW.L.Johnson.Task-oriented
collab oration withemb o diedagentsinvirtualworlds.
InJ.Cassell,J.Sullivan, S.Prevost,andE.Churchill,
editors,EmbodiedConversationalAgents.MITPress,
Cambridge,MA,2000.
[34] E.A.SchegloandH.Sacks.Op eningupclosings.
Semiotica,7:289{327,1973.
[35] C.L.Sidner.Anarticial discourselanguage for
collab orative negotiation.InProceedingsofthe
forteenthNationalConferenceoftheAmerican
AssociationforArticialIntelligence(AAAI-94),
pages814{819,1994.
[36] W.Swartout,R.Hill, J.Gratch,W.Johnson,
C.Kyriakakis,K.Lab ore,R.Lindheim, S.Marsella,
D.Miraglia, B.Mo ore,J.Morie,J.Rickel,
M.Thiebaux, L.Tuch,R.Whitney,andJ.Douglas.
Towardtheholo deck: Integratinggraphics, sound,
characterandstory.InProceedingsof5th
InternationalConferenceonAutonomousAgents,
2001.
[37] D.R.Traum.A ComputationalTheoryofGrounding
inNaturalLanguageConversation.PhDthesis,
DepartmentofComputerScience,University of
Ro chester,1994.Alsoavailable asTR545,
DepartmentofComputerScience,University of
Ro chester.
[38] D.R.TraumandJ.F.Allen. Discourseobligation s in
dialoguepro cessing. InProceedingsofthe32 nd
Annual
Meetingofthe Associationfor Computational
Linguistics,pages1{8,1994.
[39] D.R.TraumandE.A.Hinkelman.Conversation acts
intask-orientedsp okendialogue.Computational
Intelligence,8(3):575{599,1992.Sp ecial Issueon
Non-literal language.
[40] M.A.WalkerandS.Whittaker. Mixedinitiative in
dialogue: Aninvestigation into discourse
segmentation.InProceedingsACL-90,pages70{78,
1990.
[41] S.WhittakerandP.Stenton.Cuesandcontrolin
exp ert-clientdialogues. InProceedingsACL-88,pages
123{130,1988.