Learning Robust Plans for Mobile Robots from a Single Trial

Abstract

We address the problem of learning robust plans for robot navigation by observing particular robot behaviors. In this paper we present a method which can learn a robust reactive plan from a single example of a desired behavior. The system operates by translating a sequence of events arising from the effector system into a plan which represents the dependencies among such events. This method allows us to rely on the underlying stability properties of low-level behavior processes in order to produce robust plans. Since the resultant plan reproduces the original behavior of the robot at a high level, it generalizes over small environmental changes and is robust to sensor and effector noise. Introduction Recently, a number of sophisticated `reactive' planning formalisms have been developed (Firby 1989; Gat 1991; McDermott 1991; Simmons 1994), which allow a great deal of flexibility in control flow and explicitly include a notion of an intelligent plan execution system. However, the compl...