In the 1970s, some AI leaders predicted that we would soon see allmanner of artificially intelligent entities in our daily lives. Unfortu-nately, in the interim, this has been true mostly in the realm of sci-ence fiction. Recently, however, pioneering researchers have beenbringing together advances in many subfields of AI, such as robotics,computer vision, natural language and speech processing, and cogni-tive modeling, to create the first generation of robots and avatars thatillustrate the true potential of combining these technologies. The pur-pose of this article is to highlight a few of these projects and to drawsome conclusions from them for future research.

We begin with a short discussion of scope and terminology. Ourfocus here is on how robots and avatars interact with humans, ratherthan with the environment. Obviously, this cannot be a sharp dis-tinction, since humans form part of the environment for such enti-ties. However, we are interested primarily in how new interactioncapabilities enable robots and avatars to enter into new kinds of rela-tionships with humans, such as hosts, advisors, companions, andjesters.

We will not try to define robot here, but we do want to point outthat our focus is on humanoid robots (although we stretch the catego-ry a bit to include a few animallike robots that illustrate the types ofinteraction we are interested in). Industrial automation robotics,while economically very important, and a continual source ofadvances in sensor and effector technology for humanoid robots, willcontinue to be more of a behind-the-scenes contributor to our every-day lives.

The meaning of the term avatar is currently in flux. Its original andnarrowest use is to refer to the graphical representation of a person(user) in a virtual reality system. Recently, however, the required con-

Articles

SPRING 2009 29Copyright © 2009, Association for the Advancement of Arti!cial Intelligence. All rights reserved. ISSN 0738-4602

Robots and Avatars as Hosts,Advisors, Companions,

and Jesters

Charles Rich and Candace L. Sidner

! A convergence of technical progress in AI androbotics has renewed the dream of buildingartificial entities that will play significantand worthwhile roles in our human lives.We highlight the shared themes in somerecent projects aimed toward this goal.

AI Magazine, Spring 2009.

nection to a real person has been loosened and theterm avatar has been used to refer to NPCs (non-player characters) in three-dimensional computergames and to synthetic online sales representa-tives, such as Anna at ikea.com. We hope thisbroader usage will catch on and displace the termembodied conversational agent, which is somewhatconfusing, especially in the same discussion asrobots, since it is, after all, robots—not graphicalagents—that have real bodies. We will thereforeuse the term avatar in this article to refer to intelli-gent graphical agents in general.

Human Interaction Capabilities There are four key human interaction capabilitiesthat characterize the new generation of robots andavatars: engagement, emotion, collaboration, andsocial relationship. These capabilities are listedroughly in order from “low-level” (closer to thehardware and with shorter real-time constraints) to“high-level” (more cognitive), but as we will see,

there are many interdependencies among thecapabilities.

Engagement Engagement is the process by which two or moreparticipants in an interaction initiate, maintain,and terminate their perceived connection to oneanother (Sidner et al. 2005). In natural humaninteractions, engagement constitutes an intricate-ly timed physical dance with tacit rules for eachphase of an interaction.

In copresent interaction, engagement indicatorsinclude where you look, when you nod your head,when you speak, how you gesture with yourhands, how you orient your body, and how longyou wait for a response before trying to reestablishcontact. Strategies for initiating an interactioninvolve, for example, catching your potentialinterlocutor’s eye and determining whether his orher current activity is interruptible. The desire toend an interaction (terminate engagement) isoften communicated through culturally mediatedconventions involving looking, body stance (forexample, bowing), and hand gestures. Carefulempirical and computational analysis of theserules and conventions in human interaction isincreasingly making it possible for robots andavatars to connect with humans in these sameways.

Emotion There has never been any doubt about the impor-tance of emotions in human behavior, especiallyin human relationships. The past decade, however,has seen a great deal of progress in developingcomputational theories of emotion that can beapplied to building robots and avatars that interactemotionally with humans. According to the main-stream of such theories (Gratch, Marsella, and Pet-ta 2008), emotions are inextricably intertwinedwith other cognitive processing, both as ante ce -dents (emotions affect cognition) and consequen -ces (cognition affects emotions).

In terms of interacting with humans, a robot oravatar needs to both recognize the emotional stateof its human partners (through their gesture,stance, facial expression, voice intonation, and soon) and similarly express information about itsown emotional state in a form that humans canrecognize.

Collaboration Collaboration is a process in which two or moreparticipants coordinate their actions towardachieving shared goals (Grosz and Kraus 1996).Furthermore, most collaboration between humansinvolves communication, for example, to describegoals, negotiate the division of labor, monitorprogress, and so on. All the robots and avatars

Articles

30 AI MAGAZINE

Valerie.

described in this article are designed to be partici-pants in collaborations with humans (and possiblyother robots and avatars, although we focus onlyon human interactions here).

In general, collaboration is a higher-level processthat is supported by engagement; collaboration isfarther from the “hardware” and has slower real-time constraints than engagement. For example, acollaborator relies on the engagement state toknow when it is appropriate to continue with thecollaboration. However, engagement and collabo-ration are not strictly hierarchical. The state of thecollaboration can also affect how engagementbehaviors are interpreted. For example, whether ornot to interpret breaking eye contact (lookingaway) as an attempt to disengage depends onwhether the next action in the collaborationrequires looking at a shared artifact; if it does, thenbreaking eye contact does not indicate disengage-ment.

Social Relationship Most work in this area to date has involved onlyshort interactions with humans and robots oravatars (less than an hour), usually with a clearimmediate collaborative goal, such as instruction,shopping, or entertainment. Even if a user inter-acts with a system repeatedly over a long period oftime, such as return customers to a synthetic websales agent, there is typically only minor continu-ity between episodes, such as the learning of userpreferences. Furthermore, there has not generallybeen any explicit concern in the design of suchsystems towards building and maintaining long-term social relationships with humans, as wouldbe the case for similar human-human interactions.

Recently, however, as we will see shortly, sever-al researchers have begun developing robots andavatars that are designed to build and maintainsocial relationships with their users over weeks andmonths. In a sense, social relationship is the long-term correlate of engagement. The practical moti-vation for developing social relationships has beenthat the behavior change goals of these systems,such as weight loss and other better-health prac-tices, require a long time to succeed and users arenot as likely to persevere without the social rela-tionship component. Thus social relationship sup-ports collaboration, and also vice versa, since pos-itive progress toward a shared goal improves thesocial relationship.

Humanoid Robots Humanoid robots run the gamut from so-called“trash can” robots (no disrespect intended), suchas Carnegie Mellon University’s Valerie (Gockley etal. 2005) (see photo on page 30) and the NavalResearch Laboratories’ George (Kennedy et al,

2007), which simply place a face-only avatar dis-play on top of a generic mobile base, to Ishiguro’sGeminoid (Nishio, Ishiguro, and Hagita 2007),which attempts to cross the “uncanny valley”(Mori 2005) and emerge successfully on the otherside. In between are all kinds of robots withhumanlike, animallike, and cartoonlike appear-ances and dexterity in various proportions. Theapplications to which these robots are aimed are

Articles

SPRING 2009 31

George.

equally diverse. For example, the EuropeanUnion’s JAST robot (Rickert et al. 2007) mountsPhilip’s iCAT head on top of a torso with two verydexterous humanlike arms. The focus of this workis on collaborative dialogue in the domain ofassembly tasks.

Probably the most complex animallike robotconstructed to date is the Massachusetts Instituteof Technology (MIT) Media Lab’s Leonardo(Thomaz and Breazeal 2007), which has 61 degreesof freedom, 32 in the face alone. Leonardo’s ex -pres siveness is being exploited for research on therole of emotions and social behavior (thus far onlyshort-term social interaction, not building long-term social relationships) in human-robot interac-tion. The Media Lab is currently completing anequally complex, but more humanoid, robotnamed MDS (for mobile, dexterous, social), whichis roughly the size of a three-year-old child (seephoto on page 33).

Our own recent work with Mel (see photo onpage 34) (Sidner et al. 2005, 2006), a penguin wear-ing glasses with a moveable head, beak, and wings,mounted on a mobile podium base, studiedengagement behaviors in the context of what wecalled “hosting.” A robot host guides a human, orgroups of humans, around an environment (suchas a museum or a store), tells them about the envi-ronment, and supervises their interaction withobjects in the environment. Hosting is form of col-laboration and companionship aimed primarily atinformation sharing rather than long-term rela-tionship building.

Mel implemented algorithms for initiating,maintaining, and terminating engagement in spo-ken dialogues with a human collaborator. Meltracked the human’s face and gaze and, when itwas appropriate, looked at and pointed to sharedobjects relevant to the conversation. Mel also pro-duced and recognized head nods. Mel could con-verse about himself, participate in a collaborativedemonstration of a device, as well as locate a per-son in an office environment and initiate an inter-action with that person. Mel’s explicit engagementmodel included, among other things, where thehuman was currently looking and the elapsed timesince it was the human’s turn to speak. Mel alsohad explicit rules for deciding what to do when thehuman signaled a desire to disengage.

In user studies, we found that when Mel wastracking the human interlocutor’s face the humanmore often looked back at Mel when initiating adialogue turn than when Mel was not face track-ing. (Looking at your conversational partner whenyou initiate your dialogue turn is a natural behav-ior in human-human dialogues.) Furthermore,human interlocutors considered Mel more “natur-al” when he was tracking faces. Finally, humansnodded more at Mel when he recognized their

Articles

32 AI MAGAZINE

Leonardo.

JAST.

Articles

SPRING 2009 33

MDS.

Articles

34 AI MAGAZINE

Mel.

head nods and nodded back in response, as com-pared to when he did not recognize their headnods.

Kidd’s Autom (Kidd and Breazeal 2007), soon tobe commercially produced by his company, Intu-itive Automata, Inc., was designed for extended(many month) use in homes as a weight-loss advi-sor and coach. Kidd’s work builds on pioneeringresearch by Bickmore (see the next section) onlong-term social interaction and behavior changeusing avatars.

Another (at least for a time) commercially pro-duced humanoid robot is Melvin (called Reddy byits manufacturer, Robomotio, Inc.). Melvin wasdesigned by us and our colleagues at MitsubishiElectric Research Laboratories (MERL) in collabora-tion with Robomotio specifically as a cost-effectiveresearch vehicle for human-robot interaction. Hehas 15 degrees of freedom (including an expressiveface), a stereo camera, and a microphone array andspeakers and is mounted on a Pioneer mobile base.Melvin currently resides at Worcester PolytechnicInstitute (WPI) and is being used to continue theresearch on engagement and collaboration startedwith Mel described above.

Finally, a small number of researchers are tryingto develop what are called androids, that is, robots

Articles

SPRING 2009 35

Autom.

Melvin.

View color photographs of all the robots and avatars

in this articleas well as video presentations

at the author’s web site:

www.cs.wpi.edu/rich/aimag

that are ultimately indistinguishable—at least inappearance and movement—from humans. Han-son is focusing just on heads, such as Einstein(Hanson 2006), while Hiroshi Ishiguro has createdGeminoid (Nishio, Ishiguro, and Hagita 2007), afull-body android copy of himself, and Newscaster,an android copy of the well-known Japanese TVnewscaster. Ishiguro’s immediate goal for Gemi-noid is to teleoperate it as a surrogate for him inremote meetings. Unfortunately, androids are cur-rently only convincing when seated, because eventhe best biped walking robots still do not look likea natural human walking.

Limitations and Challenges Adopting the traditional decomposition of robotarchitecture into sensing, thinking, and acting, it isfair to say that the greatest barriers to achievingnatural human-robot interaction currently lie in

Articles

36 AI MAGAZINE

Einstein.

Geminoid (with Creator at Left).

Newscaster.

the sensing component (which includes interpre-tation of raw sense data into meaningful represen-tations). For the robots discussed above, this basi-cally comes down to machine vision and spokendialogue understanding.

Machine vision research has progressed signifi-cantly in recent years, notably including the devel-opment of reliable algorithms for face tracking(Viola and Jones 2001), human limb tracking(Demirdjian 2004), face recognition (Moghaddam,Jebara, and Pentland 2000) and gaze recognition(Morency, Christoudias, and Darrell 2006). Therehave also been limited improvements in objectrecognition (Torralba, Murphy, and Freeman 2004;Liebe et al. 2007), which is important for applica-tions of human-robot interaction, such as collabo-rative assembly. However, all of this technology isstill in relative infancy. For example, these algo-rithms currently perform well only when the robotitself is not moving.

These days, a kind of spoken dialogue technolo-gy is routinely used in commercial applications,

such as airline reservation telephone lines. How-ever, these systems succeed only by tightly con-trolling the conversation using system initiativeand restricted vocabularies. Unrestricted naturalconversation is beyond the capabilities of currentspoken dialogue systems, because human speechin such situations can be highly unpredictable, var-ied, and disfluent. A promising direction of currentresearch in this area is using models of dialogue toimprove speech recognition (Lemon, Gruenstein,and Peters 2002). At the current state of the art,however, human-robot interaction through spo-ken language only works when it is carefullydesigned to limit and guide the human speaker’sexpectations.

Avatars Even though some of the most difficult scientificchallenges for human-robot interaction lie in thesensing technology, this is not to say that keepingall the actuator hardware running is not a major

Articles

SPRING 2009 37

The Avatar Leonardo.

practical problem for robotics researchers, becauseit is. One can view avatars as a “solution” to thisproblem—or at least a divide-and-conquerapproach—which allows some researchers to con-centrate on the sensing and thinking components(especially regarding emotions and social relation-ship) by replacing physical actuators with graphi-cal animation and rendering technology. Thanksto the computer game and entertainment indus-

tries, very high-quality graphics and renderingtechnology is available essentially off-the-shelf.

For example, the MIT Media Lab also developeda very detailed Leonardo avatar (see photo on page37), which is substitutable for the robot. Thisapproach does, however, have some cautions.Experiments have shown (Wainer et al. 2006) thatpeople react differently overall to the physical pres-ence of a robot versus an animated character oreven viewing the same physical robot on a televi-sion screen.

Pelachaud’s Greta (2005) is a full-body avatarwith expressive gestures and facial animation,including eye and lip movements. Greta is beingused to study the expression of emotions and com-munication style in spoken dialogue, both byPelachaud and other researchers, such as Andre atthe University of Augsburg.

Cassell’s Sam (Ryokai, Vaucelle, and Cassell2002; Cassell 2004) is an example of a so-calledmixed-reality system. Sam is a virtual playmatewho is able to attend to children’s stories and tellthem relevant stories in return. Furthermore, chil-dren can pass figurines back and forth from thereal to the virtual world. Sam has been demon-strated to improve children’s literacy skills.

Bickmore’s Laura (Bickmore and Picard 2005) isaimed toward the same class of applications later

Articles

38 AI MAGAZINE

Greta.

Laura.

Sam.

addressed by Autom, namely health behaviorchange (for example, diet and exercise). LikeAutom, Laura was designed to develop long-termsocial relationships and is the first such avatar tohave a month-long daily relationship with a user,in fact, several users. For example, people gettingexercise advice from Laura over a several week spanwere shown to be responding to her sociallyaccording to standard psychological measures.Bickmore’s more recent research includes a pilotstudy at the Boston Medical Center GeriatricAmbulatory Practice, which showed that patientsusing Laura as an exercise coach daily for twomonths walked significantly more compared to acontrol group (Bickmore et al. 2005).

The University of Southern California’s Insti-tute for Creative Technologies (ICT) is developinga collection of realistic soldier and civilian avatarsto inhabit virtual worlds for interactive militarytraining (Swartout et al. 2005). For example, Sgt.Blackwell (Leuski et al. 2006) is a wisecracking mil-itary character who answers questions about thearmy and technology. Among other things, thiswork is pushing the boundaries of spoken dia-logue technology.

Conclusions Judging from these recent projects, the two areaswhere robots and avatars are soonest likely to havea significant and worthwhile role in our lives arehealth and education/training. Autom, Sam, Lau-ra, and Sgt. Blackwell are indicators of what toexpect in these areas.

Closely related to this cluster are applicationsthat can be generally characterized as assistive,either socially or physically. For example, Feil-Seiferand Mataric (2005) have developed a robotic play-mate (reminiscent of the Sam avatar) for autisticchildren. In this work, the real purpose of the inter-action is to teach social skills; the human-robot col-laborative task is only a means to that end.

Obviously, as robots become able to use theirhands and arms safely in close proximity tohumans, many physically assistive applications,such as helping the elderly, will become feasible.Furthermore, as compared to the partially compet-ing approach of ubiquitous computing, in whichthe entire environment is instrumented and auto-mated, a humanoid robot can also offer compan-ionship (emotion and social relationship). Evi-dence already suggests that people respondpositively to such robots.

Of course, the “killer app” is to add domestic ser-vant to the list of roles in the title of this article.Although many researchers have this goal inmind, a general-purpose domestic robot, able towork in an uncontrolled home environment, isstill a long ways off.

Almost all of the interaction between humansand avatars or robots thus far have been one-to-one (dialogues). Clearly, however, robots workingin human-populated environments will need to be

Articles

SPRING 2009 39

Sgt. Blackwell.

able to carry on multiparty conversations. The col-laborative model underlying such conversations isreasonably well understood, for example, by Groszand Kraus (1996), but the engagement aspects havebeen much less studied. ICT has developed a pio-neering system in which a human trainee engagesin a delicate wartime negotiation with two avatarsrepresenting a village doctor and elder (Traum etal. 2008). Matsusaka (2005) has done importantinitial work on gaze in three-party conversations,which he implemented for avatars at ICT and laterfor the Mel robot at MERL.

Overall, research on interacting with robots andavatars is valuable not only for its applications butalso for its contributions to understanding humanbehavior. For example, our research on engage-ment for Mel started with detailed analysis of theengagement behaviors in videotaped human-human interactions. Similarly, Bickmore’s Laurahas served as a research platform for studyinghuman social dialogue, as well as being a practicalaid for helping people change their diet and exer-cise habits.

Returning finally to the four key human inter-action capabilities discussed at the start of this arti-cle, we would like to emphasize emotion and socialrelationship as the current research frontier. We arejust beginning to understand how make thesecapabilities (including even humor—see the danceroutine in the Melvin video) part of the systems webuild. The next decade in this field will undoubt-edly prove to be challenging and intriguing!

AcknowledgementsThis article is based on a invited talk by C. Sidnerat the Twenty-First International Florida ArtificialIntelligence Society Conference (FLAIRS’08),Coconut Grove, FL, May 2008.

References Bickmore, T.; Caruso, L.; Clough-Gorr, K.; and Heeren, T.2005. It’s Just Like You Talk to a Friend: Relational Agentsfor Older Adults. Interacting with Computers 17(6): 711–735. Bickmore, T,. and Picard, R. 2005. Establishing and Main-taining Long-Term Human-Computer Relationships.ACM Transactions on Computer Human Interaction (ToCHI)59(1): 21–30. Cassell, J. 2004. Towards a Model of Technology and Lit-eracy Development: Story Listening Systems. Journal ofApplied Developmental Psychology 25(1): 75–105. Demirdjian, D. 2004. Combining Geometric- and View-Based Approaches for Articulated Pose Estimation. InComputer Vision—ECCV 2004: 8th European Conference onComputer Vision. Lecture Notes in Computer Science Vol-umes 3021–4. Berlin: Springer.Feil-Seifer, D. J., and Mataric, M. J. 2005. Defining Social-ly Assistive Robotics. In IEEE 9th International Conferenceon Rehabilitation Robotics: Frontiers of the Human-MachineInterface, 465–468. Piscataway, NJ: Institute of Electricaland Electronics Engineers.Gockley, R.; Bruce, A.; Forlizzi, J.; Michalowski, M.;Mundell, A.; Rosenthal, S.; Sellner, B.; Simmons, R.;Snipes, K.; Schultz, A.C.; and Wang, J. 2005. Designing

Articles

40 AI MAGAZINE

A Village Doctor and Elder.

Robots for Long-Term Social Interaction. InProceedings, International Conference on Intel-ligent Robots and Systems. Piscataway, NJ:Institute of Electrical and Electronics Engi-neers. Gratch, J.; Marsella, S.; and Petta, P. 2008.Modeling the Cognitive Antecedents andConsequents of Emotion. Cognitive SystemsResearch 10(1): 1–5. Grosz, B. J., and Kraus, S. 1996. Collabora-tive Plans for Complex Group Action.Artificial Intelligence 86(2): 269–357. Hanson, D. 2006. Expanding the DesignDomain of Humanoid Robots. Paper pre-sented at the 5th International Conferenceof the Cognitive Science, special session onAndroid Science, Vancouver, BC, 26 July. Kennedy, W. G.; Bugajska, M.; Marge, M.;Adams, W.; Fransen, B. R.; Perzanowski, D.;Schultz, A. C.; and Trafton, J. G. 2007. Spa-tial Representation and Reasoning forHuman-Robot Collaboration. In Proceedingsof the Twenty-Second Conference on ArtificialIntelligence. 1554–1559. Menlo Park, CA:AAAI Press. Kidd C. D., and Breazeal, C. 2007. A Robot-ic Weight Loss Coach. In Proceedings of theTwenty-Second Conference on Artificial Intelli-gence. Menlo Park, CA: AAAI Press. Leibe, B.; Cornelis, N.; Cornelis, K.; and vanGool, L. 2007. Dynamic 3D Scene Analysisfrom a Moving Vehicle. In Proceedings of the2007 IEEE Computer Society Conference onComputer Vision and Pattern Recognition. Pis-cataway, NJ: Institute of Electrical and Elec-tronics Engineers. Leuski, A.; Patel, R.; Traum, D.; andKennedy, B. 2006. How to Talk to a Holo-gram. In Proceedings of the 2006 Internation-al Conference on Intelligent User Interfaces.New York: ACM Press. Lemon, O.; Gruenstein, E.; and Peters, S.2002. Collaborative Activities and Multi-Tasking In Dialogue Systems. TraitementAutomatique des Langues (TAL), Special issueon Dialogue, 43(2): 131-154. Matsusaka, Y. 2005. Recognition of 3rd Par-ty Conversation Using Prosody and Gaze.In Interspeech 2005: Proceedings of the 9thEuropean Conference on Speech Communica-tion and Technology, 1205–1208. Grenoble,France: International Speech Communica-tion Association.Moghaddam, B.; Jebara, T.; and Pentland,A. 2000. Bayesian Face Recognition. PatternRecognition 33(11): 1771–1782. Morency, L.-P.; Christoudias, C. M.; andDarrell, T. 2006. Recognizing Gaze Aver-sion Gestures in Embodied ConversationalDiscourse. In Proceedings of the Eighth Inter-national Conference on Multimodal Interac-tions. New York: Association for Comput-ing Machinery.

Mori, M. 2005. On the Uncanny Valley.Paper presented at the Humanoids 2005Workshop: Views of the Uncanny Valley,Tsukuba, Japan, 5 December. Nishio, S.; Ishiguro, H.; and Hagita, N.2007. Can a Teleoperated Android Repre-sent Personal Presence? A Case Study WithChildren. Psychologia 50(4): 330–343. Pelachaud, C. 2005. Multimodal ExpressiveEmbodied Conversational Agent. Paperpresented at the 13th Annual ACM Inter-national Conference on Multimedia, BraveNew Topics Session, Singapore, November6–11. Rickert, M.; Foster, M. E.; Giuliani, M.; By,T.; Panin, G.; and Knoll, A. 2007. Integrat-ing Language, Vision and Action forHuman Robot Dialog Systems. In Proceed-ings of the 8th International Conference onHuman-Computer Interaction, 987–995.Hillsdale, NJ: Lawrence Erlbaum AssociatesInc. Ryokai, K.; Vaucelle, C.; and Cassell, J.2002. Literacy Learning by Storytellingwith a Virtual Peer. In Proceedings of Com-puter Support for Collaborative Learning.Mahwah, NJ: Lawrence Erlbaum Associates.Sidner, C.; Lee, C.; Kidd, C.; Lesh, N.; andRich, C. 2005. Explorations in Engagementfor Humans and Robots. Artificial Intelli-gence 166(1–2): 140–164. Sidner, C.; Lee, C.; Morency, C.; Forlines, C.2006. The Effect of Head-Nod Recognitionin Human-Robot Conversation. In Proceed-ings of the 2006 ACM Conference on HumanRobot Interaction, 290–296. New York: Asso-ciation for Computing Machinery,Swartout, W.; Gratch, J.; Hill, R.; Hovy, E.;Lindheim, R.; Marsella, S.; Rickel, J.; andTraum, D. 2005. Simulation Meets Holly-wood: Integrating Graphics, Sound, Storyand Character for Immersive Simulation.In Multimodal Intelligent Information Presen-tation, ed. O. Stock and M. Zancanaro.Berlin: Springer. Thomaz, A. L., and Breazeal, C. 2007. RobotLearning via Socially Guided Exploration.In IEEE 6th International Conference onDevelopment and Learning. Piscataway, NJ:Institute of Electrical and Electronics Engi-neers.Torralba, A.; Murphy, K. P.; and Freeman,W. T. 2004. Sharing Features: EfficientBoosting Procedures for Multiclass ObjectDetection. In Proceedings of the 2004 IEEEComputer Society Conference on ComputerVision and Pattern Recognition (CVPR), 762–769. Piscataway, NJ: Institute of Electricaland Electronics Engineers.Traum, D.; Marsella, S.; Gratch, C.; Lee, J.;and Harthold, A. 2008. Multi-Party, Multi-issue, Multi-Strategy Negotiation for Multi-Modal Virtual Agents. In Intelligent Virtual

Articles

SPRING 2009 41

Agents: Proceedings of the 8th InternationalConference. Lecture Arti!cial Intelligence, Vol-ume"5208. Berlin: Springer.Viola, P., and Jones, M. 2001. Rapid ObjectDetection Using a Boosted Cascade of Sim-ple Features. In Proceedings of the IEEE Con-ference on Computer Vision and Pattern Recog-nition, 905–910. Piscataway, NJ: Institute ofElectrical and Electronics Engineers. Wainer, J.; Feil-Seifer, D. J.; Shell, D. A.; andMataric, M. J. 2006. The Role of PhysicalEmbodiment in Human-Robot Interaction.In IEEE Proceedings of the International Work-shop on Robot and Human Interactive Com-munication (RO-MAN), 117–122. Piscat-away, NJ: Institute of Electrical andElectronics Engineers.

Charles Rich is a professor of computer sci-ence and a member of the InteractiveMedia and Game Development faculty atWorcester Polytechnic Institute. He waspreviously a distinguished research scien-tist and founding member of MitsubishiElectric Research Laboratories. Rich earnedhis Ph.D. at the MIT Artificial IntelligenceLaboratory, where he was a founder anddirector of the Programmer’s Apprenticeproject. He is a Fellow and past councilor ofAAAI and a senior member of IEEE. He hasserved as chair of the 1992 InternationalConference on Principles of KnowledgeRepresentation and Reasoning, cochair ofthe 1998 National Conference on ArtificialIntelligence, and program cochair of the2004 and chair of the 2010 InternationalConference on Intelligent User Interfaces.His e-mail address is rich@wpi.edu.

Candace L. Sidner is a division scientist atBAE Systems Advanced Information Tech-nology and a consultant to Tim Bickmore’sbehavior change dialogue project at North-eastern University and Rich's human-robotinteraction project at WPI. She earned herPh.D. at the MIT Artificial Intelligence Lab-oratory. Sidner is a Fellow and past coun-cilor of AAAI, an associate editor of Arti!-cial Intelligence, and a senior member ofIEEE. She has served as president of theAssociation for Computational Linguistics,chair of the 2001 and program cochair ofthe 2006 International Conference onIntelligent User Interfaces (IUI), cochair ofthe 2004 SIGDIAL Workshop on Discourseand Dialogue, chair of the 2007 NAACLHuman Language Technology (NAACL-HLT) conference, and as a member of thescienti!c advisory boards for the E.U. Cog-nitive Systems for Cognitive Assistants(CoSy) project, SIGDIAL, IUI and NAACL-HLT. Her e-mail address is candy.sidner@alum.mit.edu.