How to move (physically speaking) in a multi-agent world (abstract)

Abstract

We consider the case of a physical agent (e.g., a robot) who accomplishes tasks by moving in a physical world populated by other physical agents. We discuss various approaches for planning the motions of such an agent. These approaches range between two extremes. At one extreme, the agent is a super-agent that plans the coordinated motions of all the other agents. At the other extreme, the agent deals with the other agents as if they were unexpected moving obstacles detected through sensors during the execution of motion plans. Between these two extremes, many other approaches are possible by engineering and combining basic motion planning methods (and higher-level reasoning methods) in different ways. These approaches differ in the amont of prior knowledge and the nature of inter-agent communication they require, as well as in their completeness, computational complexity, and their tolerance to unexpectedness. If a truly autonomous agent in a multi-agent world is to be eventually built, it is very likely that it will have to embed several of these approaches.