T8 - FULL DAY ================================================================= Teamwork among Agents, Humans and Robots: Theory and Practice Barbara Grosz, Charlie Ortiz and Milind Tambe ==================================================================== ABSTRACT OF TUTORIAL Teamwork is a fundamental area of multiagents research and important for a wide range of applications. Clearly, a significant number agent teams applications have already emerged, such as synthetic agent teams in virtual environments for training, software assistant teams to support human organizations, sensor teams for target tracking, mobile robot teams for security and mine-clearance, robot-agent-person teams for disaster rescue. Indeed, as we look towards the future, it is difficult to think of real- world applications where an agent can act alone. Instead, most applications will depend on an agent's teamwork or collaboration with other agents, including robots, softbots and people appears critical in almost every real-world application. This tutorial will survey the state of the art of both of the theory of multiagent teams and practical implementations of teamwork-capable agent systems. We will discuss two different threads of research on teamwork theory. The first thread, based on belief-desire-intention (BDI) models, will focus on theories such as the joint intentions theory (Cohen and Levesque), SharedPlans theory (Grosz and Kraus) and planned team activity theory (Kinny et al). The second thread, based on economic team theory (Marschak and Radner), provides a decision-theoretic perspective. We will cover this thread, and its recent reincarnation in dynamic domains in the form of decentralized partially observable markov decision processes (Decentralized POMDP) family of models for teamwork and cooperative multiagent systems. We will discuss applications of such models in analysis of multiagent systems. We will also devote half of the tutorial to practical systems constructed by exploiting teamwork theory as the basis. Indeed, one key lesson learned in practical systems is that in complex, dynamic environments, creating fixed and domain-specific coordination plans is highly problematic: these plans are not reusable across domains and their lack of flexibility can lead to severe failures. Instead, a new approach based on developing general teamwork models, i.e., domain-independent, reusable team coordination algorithms, appears to provide more promise. We will discuss research on general team coordination algorithms developed by Tambe, Jennings, and several others, and cover applications such as teams of synthetic pilots, simulated soccer players, software assistant teams and robot-agent-person teams developed using these coordination algorithms. The last part of the tutorial will cover agent-human teamwork, and key issues in such teamwork, in particular, mixed-initiative and adjustable autonomy. In addition to the theoretical aspect of these key new issues, we will provide an overview of research systems as well as industrial systems, including work at Microsoft and MERL, which involve agent-human teams. =============================================================================== BACKGROUND KNOWLEDGE REQUIRED The potential target is graduate students and interested researchers who have some familiarity with agents and have already taken an introductory course in Artificial Intelligence. The tutorial will be particularly relevant to those researchers who are working in the area of cooperative multiagent systems, and agent-human interaction systems. Other senior researchers in the field who are interested in brushing up on the latest in the teamwork arena are also welcome. ================================================================================= INTENDED AUDIENCE Teamwork is a critical topic of research in multiagent environments. Significant numbers of papers in AAMAS'2002 were focused on teamwork. A similar tutorial at AAMAS'2002 was deemed to be highly successful, attracting one of the largest number of attendees in any tutorial at AAMAS'2002. Given the breadth of our presentation, which will include agent, human and robotic teamwork, we believe that significant number of attendees will find our tutorial of interest. ============================================================================= DETAILED OUTLINE This tutorial will be divided into four logical parts as outlined below. We will encourage a great deal of interaction with the students, rather than just going through the material at a rapid pace: Introduction, Motivation and overview: While we will assume that attendees are familiar with the notion of agents, we will quickly (re)introduce the notion of agents, and follow that up with a discussion and definition of teamwork. Teamwork is not just a union of simultaneous coordinated activity, but rather involves a common goal and more. Introduce several key agent-agent and agent-human team applications, including videos of some of these applications. Outline key challenges in modeling teamwork and reasoning about team participation and explain the major challenge of team coordination in complex, uncertain domains and the need for domain-independent team coordination algorithms. Teamwork theory: Theory will be covered in two sections. The first section will examine leading belief-desire-intention teamwork theories, and in particular the SharedPlans (Grosz/Kraus/Sidner) theory of teamwork, as well as the joint intentions theory and their differences. Provide an overview of recent advances in these theories (e.g., from ICMAS'2000). The second section will discuss economic Team theory (Marschak and Radner), and in particular, new models related to this research, such as the decentralized POMDP models, their complexity results, and their application to analyses of multiagent systems (Boutilier, Kaebling, Lesser, Pynadath/Tambe, and others). Practical agent teams based on theory: Explain practical implementations of theory, e.g., the STEAM teamwork model(Tambe) that addresses domain-independent team coordination. Explain weaknesses in teamwork theory and how practical systems address them in their domains. Cover practical applications of teamwork models, such as STEAM's application in helicopter teams in battlefield simulations and in robot-agent-person teams. We will also cover other teamwork-theory based architectures. Finally, we will provide pointers to teamwork-related software packages and explain how these packages could be used in building applications. We will also provide demonstrations of agent teamwork systems developed --- via movies where necessary or practical illustrations where possible. Agent-human mixed teams: As we focus on heterogeneous teams, agent-human teams bring up a key challenge of mixed-initiative or adjustable autonomy. Agents must flexibly share initiative with teammates (humans), dynamically varying own autonomy to give decision-making control to humans. Provide an overview of approaches to adjustable autonomy, including work at Microsoft (Horvitz), MERL (Collagen), Electric-Elves work at USC, which involve agent-human teams and others. We will also cover work at Harvard University on DIAL and Writer's aid, and some of the related work on human-computer interfaces. ============================================================================ BIOGRAPHIES OF PRESENTERS Barbara J. Grosz Higgins Professor of Natural Sciences Division of Engineering and Applied Sciences Harvard University Cambridge, Massachusetts 02138, USA grosz@eecs.harvard.edu (617) 495-3673, office Professor Grosz's research is addressing fundamental problems in modeling collaborative activity and is developing computer systems ("agents") able to collaborate with each other and their users. The SharedPlans model of collaboration (developed with S. Kraus) provides a specification for the construction of such collaborative agents. Prof. Grosz is extending this model and using it to construct computer agents that work together in teams. Professor Grosz also investigates the basic structures and processes by which people use natural languages to communicate. She has developed (with C. Sidner) a theory of discourse structure that specifies how discourse interpretation depends on interactions among speaker intentions, attentional state, and linguistic form. These two strands of research are being combined in an effort that aims to provide the scientific and technological base for a new paradigm for human-computer interaction, one that enables the principled design of multi-modal dialogue-supporting interfaces. Honors and Professional Society Participation: Fellow, American Association for the Advancement of Science; Fellow, American Association for Artificial Intelligence; Distinguished Alumna Award in Computer Sciences and Engineering, University of California at Berkeley, 1997. Charles L. Ortiz, Jr. Artificial Intelligence Center SRI International Room EJ278 333 Ravenswood Ave Menlo Park, CA 94025 ortiz@ai.sri.com (650) 859-4461, office (650) 859-3735, fax Dr. Ortiz is the Program Manager of the Collaboration Science and Technology Program in the AI Center of SRI International. He received an S.B.in physics from MIT and an M.S.in computer science from Columbia University. His Ph.D. in computer and information science from the University of Pennsylvania in 1996 was for his work on reasoning about action, counterfactuals, and causation. From 1996-1998 he worked on collaborative planning systems and the design of rational agent architectures as a postdoctoral fellow at Harvard University. At SRI he is the PI for several projects on multiagent collaboration. As a postdoctoral fellow at Harvard, he conducted research on collaborative planning systems and rational agency. His research interests include commonsense reasoning (action, causation, counterfactuals) as well as collaborative planning and rational agent architectures. Milind Tambe Associate Professor, Computer Science Project Leader, Information Sciences Institute Computer Science Dept and Information Sciences Institute University of Southern California Henry Salvatori Computer Center 232 Los Angeles, CA 90089-0781 Tel: 213-740-6447 Fax: 213 740 7285 http://www.isi.edu/teamcore/tambe email: tambe@usc.edu Milind Tambe obtained his PhD from the School of Computer Science at Carnegie Mellon University in 1991. He is currently an Associate Professor of Computer Science at the University of Southern California and a project leader at USC's Information Sciences Institute (ISI). His research interests are in areas of multiagent systems, particularly in topics such as teamwork, coordination, negotiations and adjustable autonomy. He will be the general co-chair for the International Conference on Autonomous Agents and Multiagent Systems (AAMAS) in 2004, and has served as the program co-chair for the International conference on multiagent systems (ICMAS) in 2000. He is an associate editor for the Journal for AI research (JAIR) and he is on the editorial board of the Journal of Autonomous Agents and Multiagent Systems (JAAMAS), and IEEE Intelligent Systems. He was also a chair of the organizing committee for the first Americas' Agents School, and he is currently a member of the board of directors of the International Foundation for Multiagent Systems and a member of the steering committee for the Agents, Theories, Architectures and Languages (ATAL). He is also a former trustee of the RoboCup Federation.