Collaborative Discourse, Engagement and Always-On Relational Agents

We summarize our past, present and future research related to human-robot dialogue, starting with its foundations in collaborative discourse theory, continuing to our current research on recognizing and generating engagement, and concluding with an outline of new work we are beginning on the modeling of long-term relationships between humans and robots. This paper is a summary of our past, present and future research related to human-robot dialogue. In the first section below, we describe the main elements of collaborative discourse theory and the architecture of a collaborative interaction manager, which serve as the foundation of all of our later work. In the second section, we discuss current work on engagement, which underlies and supports collaboration and dialogue. Finally, in the third section, we outline new research we are beginning, which focuses on robots (and other agents) that are “always on” and therefore need to build long-term relationships with humans in their environment. More details on each of these topics can be obtained by referring to the cited publications. Collaborative Discourse The common thread through almost all of our research for the past many years has been the view that, whenever there is interaction—and especially communication—between two intelligent agents, collaborative discourse theory provides important insights into what is going on and helpful guidance in designing computer tools to support the interaction. Human-robot dialogue is the most recent example of where we have applied this approach. Collaboration is a process in which two or more participants coordinate their actions toward achieving shared goals. Most collaboration between humans involves communication. Discourse is a technical term for an extended communication between two or more participants in a shared context, such as a collaboration. Collaborative discourse theory thus refers to a body of empirical and computational research about how people communicate in the context of a collaboration. Copyright c © 2010, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. SharedPlans Grosz and Sidner’s SharedPlans (Grosz and Sidner 1986; 1990; Grosz and Kraus 1996) provide a general computational model of how collaborative, coordinated activity emerges from the individual beliefs and intentions (goals and plans) of the collaborators. Two collaborators have a SharedPlan for a given goal when they mutually believe that: (a) they have a common (shared) goal; (b) they have agreed on a recipe to accomplish the goal; (c) they are each capable of performing their respective actions; (d) each intends to perform their respective actions; and (e) they are both committed to the overall success of the collaboration. In a typical collaboration, not all of the five conditions above are satisfied at the start. Instead, SharedPlans typically start in a partial state, e.g., having a shared goal, and incremental refinement of the SharedPlan is interleaved with performing actions that contribute toward the goal. Along the way, much of the communication between the collaborators often has to do with refining the SharedPlan, such as negotiating who should do which action. Collaborative Interaction Manager We have implemented two collaborative interaction managers based on SharedPlans, Collagen (Rich and Sidner 1998; Rich, Sidner, and Lesh 2001) and its recent successor, Disco. Collagen has been used to build more than a dozen human-computer collaborative systems. Unlike most so-called “dialogue managers,” Collagen and Disco manage both the conversational and the task structure of an interaction. This is because, according to SharedPlan theory, these two structures are deeply intertwined. Both Collagen and Disco currently support only two-participant discourse (dialogue), although we have experimented with multi-participant extensions to both of them. Disco differs from Collagen mainly in using the ANSI/CEA-2018 standard, whose development was led by Rich (2009), for representing task models. Disco also lacks the logical inference and truth maintenance facilities included in Collagen. Disco is written in Java and is distributed under the MIT open-source license; a copy may be obtained by sending email to Rich. Figure 1 shows the architecture of both Collagen and Disco. The two key data structures in this architecture are the task model and the discourse state. The task model is To appear in Dialog with Robots, AAAI Fall Symposium, Washington, DC, November 2010.