The Current State of AI Planning Research

Planning is designing the behavior of an agent. The classical theory of planning assumes that the planner has perfect information, and needs to bring about a deenite result, by nding a series of actions that brings it about. These classical assumptions have been questioned, but are quite reasonable for many engineering applications. The problem is that traditional algorithms, while elegant, are not eeective on realistic-sized problems. Recently some more promising algorithms have been developed, based on avoiding costly searches by doing a more exhaustive analysis at each search state. However, these algorithms may still not be ready for the kinds of problems that arise in manufacturing and other domains where real geometry and kinematics are important.