Plan Failure Analysis: Formalization and Application in Interactive Planning Through Natural Language Communication

Chitta Baral, Tran Cao Son, Michael Gelfond, Arindam Mitra

Abstract: While most robots in human robot interaction scenarios take instructions from humans, the ideal would be that humans and robots collaborate with each other. The Defense Advanced Research Projects Agency Communicating with Computer program proposes the collaborative blocks world scenario as a testbed for this. This scenario requires the human and the computer to communicate through natural language to build structures out of toy blocks. To formulate and address this, we identify two main tasks. The first task, called the plan failure analysis, demands the robot to analyze the feasibility of a task and to determine the reasons(s) in case the task is not doable. The second task focuses on the ability of the robot to understand communications via natural language. We discuss potential solutions to both problems and present prototypical architecture for the integration of planning failure analysis and natural language communication into an intelligent agent architecture.