Learning momentum: online performance enhancement for reactivesystems

摘要

The authors describe a reactive robotic control system whichincorporates aspects of machine learning to improve the system's abilityto navigate successfully in unfamiliar environments. This systemovercomes limitations of completely reactive systems by exercisingonline performance enhancement without the need for high-level planning.The goal of the learning system is to give the autonomous robot theability to adjust the scheme control parameters in an unstructureddynamic environment. The results of a successful implementation thatlearns to navigate out of a box canyon are presented. This system neverresorts to a high-level planner, but instead learns continuously byadjusting gains based on the progress made so far. The system issuccessful because it is able to improve its performance in reaching agoal in a previously unfamiliar and dynamic world