Manuela M. Veloso
School of Computer Science Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA 15213
To appear in the 1999 AAAI Spring Symposium on
Search Techniques for Problem Solving under
Uncertainty and Incomplete Information
Most real world environments are hard to model completely and correctly,
especially to model the dynamics of the environment. In this paper we present
our work to improve a domain model through learning from execution, thereby
improving a task planner's performance. Our system collects execution traces
from the robot, and automatically extracts relevant information to improve the
domain model. We introduce the concept of {\em situation-dependent rules},
where situational features are used to identify the conditions that affect
action achievability. The system then converts this execution knowledge into a
symbolic representation that the planner can use to generate plans appropriate
for given situations.