Collaborative Plans for Group Activities* Barbara Grosz Division of Applied Sciences Harvard University Cambridge, MA 02138 USA grosz@das.harvard.edu Abstract The original formulation of SharedPlans [Grosz and Sidner, 1990] was developed to provide a model of collaborative planning in which it was not necessary for one agent to have in- tentions toward an act of a different agent. This formulation provided for two agents to coordinate their activities without introducing any notion of jointly held intentions (or, 'we- intentions'). However, it only treated activities that directly decomposed into single agents ac- tions. In this paper we provide a revised and expanded version of SharedPlans that accom- modates actions involving groups of agents as well as complex actions that decompose into multi-agent actions. The new definitions also allow for contracting out certain actions, and provide a model with the features required in Bratman's account of shared cooperative activ- ity [Bratman, 1992]. A reformulation of the model of individual plans that meshes with the definition of SharedPlans is also provided. 1 Introduction Collaboration in planning and acting is an essential in- gredient of multi-agent cooperative problem solving. In this paper we present a model of collaborative planning that supports cooperative problem solving by teams con- sisting of humans and computer systems. The model deals more completely with collaboration than previous theories did in meeting two criteria. First, collabora- tive planning and activity cannot be analyzed simply in terms of the plans of individual agents, but require an integrated treatment of the beliefs and intentions of the different collaborating agents. Second, collaborative planning is a refinement process; a partial plan descrip- tion is modified over the course of planning by the mul- * We thank Joyce Friedman for many provoking questions, and Karen Lochbaum for the same and for helpful comments on many drafts. This research was initiated when the first author was a Harold Perlman Visiting Professor, Hebrew Uni- versity, Jerusalem. Partial support for the first author was provided by U S WEST Advanced Technologies. The second author was supported by NSF Grant No. IRI-9123460. Sarit Kraus Dept. of Mathematics and Computer Science Bar Ilan University, Ramat Gan, 52900 Israel sarit@bimacs.cs.biu.ac.il tiple collaborating agents. This model grew out of an attempt to provide an adequate treatment of the in- tentional component of discourse structure [Grosz and Sidner, 1986]. However, many multi-agent situations re- quire that agents have an ability to plan and act to- gether; merely avoiding conflicting actions or situations is not sufficient. Thus, the model is applicable not only to natural language processing, but also to the general problem of the design of computer-based collaborating agents. The original formulation of the SharedPlan model of collaborative planning [Grosz and Sidner, 1990] extended Pollack's mental state model of plans [Pollack, 1990] to the situation in which two agents jointly have a plan to perform some action requiring actions by both agents. Pollack s definition of the individual plan of an individ- ual agent to do an action a includes four constituent mental attitudes: (1) belief that performance of certain actions would entail performance of we will refer to the as constituting "a recipe for " (2) belief that the agent could perform each of the (3) intentions to do each of the (4) an intention to do a by do- ing the To define SharedPlans, Grosz and Sidner modified these components to provide a specification of the set of beliefs and intentions required for collabora- tive action. In subsequent work [Lochbaum et al., 1990; Lochbaum, 1991], algorithms were provided for con- structing and augmenting SharedPlans in the context of a dialogue. Although this formulation overcame several problems of previous models of planning for discourse (e.g. the treatment of intentions of one agent toward another agents actions in applications of speech act theory [Allen and Perrault, 1980]), it had several problems that emerged when we attempted to apply it [Lochbaum et al., 1990; Lochbaum, 1991]. First, the original model presumed that every multi-agent action decomposed di- rectly into single agent actions, a similar assumption underlies several alternative models (e.g. [Cohen and Levesque, 1990]). As a result, the model did not ad- equately provide for complex activities involving joint activity at multiple levels or for meshing of individual plans for individual action with collaborative plans for joint activity. Second, the model did not account for the commitment of an agent to the success of a collaborative partner's actions. This omission combined with the first Grosz and Kraus 367