Becoming Increasingly Reactive

Abstract

augmenting its reactive component whenever it is forced to We describe a robot control architecture which plan. When used to control a laboratory mobile robot, the combines a stimulus-response subsystem for rapid Theo-Agent in simple cases learns to reduce its reaction reaction, with a search-based planner for handling time for new tasks from several minutes to less than a unanticipated situations. The robot agent continually second. chooses which action it is to perform, using the stimulus- The research reported here is part of our larger effort response subsystem when possible, and falling back on the toward developing a general-purpose learning robot planning subsystem when necessary. Whenever it is architecture, and builds on earlier work described in forced to plan, it applies an explanation-based learning [Blythe and Mitchell 89]. We believe that in order to mechanism to formulate a new stimulus-response rule to become increasingly successful, a learning robot will have