Ron Petrick's Webpages

Position statement talk

Knowledge-Level Planning for Human-Robot Interaction, R. Petrick, slides from a position statement talk presented at the Dagstuhl Seminar on Planning and Robotics (17031), Schloss Dagstuhl, Leibniz-Zentrum für Informatik, Dagstuhl, Germany, 2017-01-16.

[ slides ]


At a basic level, the automated planning problem is one of context-dependent action selection: given an initial state, a domain description, and a set of goals, generate a sequence of actions whose execution will bring about the goal conditions. However, the problem of action selection is not unique to automated planning. One important field where this issue is also of primary concern is that of spoken dialogue systems, whose tools play a central role in addressing the problem of human-robot interaction. At the heart of the dialogue system is the interaction manager whose primary task is to carry out a form of action selection: based on the current state of an interaction, the interaction manager makes a high-level decision as to which spoken, non-verbal, and task-based actions the system should apply. An important aspect of research in this area has been the development of toolkits to support the construction of end-to-end systems. Given the parallels between the planning and dialogue tasks, our recent work has explored the application of automated planning techniques to human-robot interaction (HRI) as an alternative to standard dialogue system toolkits (such as Trindikit, COLLAGEN, IrisTK, OpenDial, among others).

While the link between natural language processing and automated planning has a long tradition, going back to at least the 1980s, in recent years the two communities have focused on different problems and solutions, with planning for natural language problems largely overlooked in favour of more special-purpose solutions. For instance, the interactive systems toolkits attempt to offer a one-stop solution for system building combining action selection, representation, and technical architectures. In contrast, the planning community has focused on defining domains in common representation languages like PDDL and comparing different domain-independent strategies within this context through events like the International Planning Competitions; the study of the representation languages themselves has also led to a better understanding of the trade-offs between different representations.

Our own work in this area has focused on applying domain-independent knowledge-level planning techniques to the problem of action selection in human-robot interaction. In particular, the beliefs of the planning agent (robot) about the world and other agents are represented, and sensing actions are used to model certain types of information-gathering speech acts. Task-based actions are also planned using the same general-purpose planning mechanisms.

However, the problem of human-robot interaction also offers some wider opportunities and lessons for the planning community. First, the presence of action selection at the core of interaction management offers the obvious possibility of applying other types of planning techniques. Second, the nature of the applications addressed by many HRI systems also highlights the importance of building real-world systems---an area that has gained wider traction in the planning community but one that is still somewhat outside the mainstream of most planning research. Finally, the process for evaluating robot-based dialogue systems, and in particular the role of human users, also presents new directions and challenges for planning.