Learning Situation-Dependent Rules:
Improving Planning from Robot Execution

Karen Zita Haigh and Manuela M. Veloso,

School of Computer Science, Carnegie Mellon University

Submitted to AAAI '98

Real world robot tasks are so complex that it is hard to hand-tune all of the domain knowledge, especially to model the dynamics of the environment. In this paper we present techniques for learning from real-world robot execution to improve planning and execution performance. We present our work learning from execution to improve a task planner's performance. Our system collects execution traces from the robot, and automatically extracts relevant information to improve the efficiency of generated plans. We introduce the concept of situation-dependent rules, where situational features are used to identify the conditions that affect action achievability. It converts this execution knowledge into a sybmolic representation that the planner can use to generate plans appropriate for given situations.

  • Compressed Postscript (7 pages, 480K)