Behavior Fusion Estimation for Robot Learning from Demonstration

Abstract

A critical challenge in designing robot systems that learn from demonstration is the ability to map the behavior of the trainer as sensed by the robot onto an existing repertoire of the robot's basic/primitive capabilities. Observed behavior of the teacher may constitute a combination (or superposition) of the robot's individual primitives. Once a task is demonstrated, our method learns a fusion (superposition) of primitives (as a vector of weights) applicable to situations encountered by the robot for performing the same task. Our method allows a robot to infer essential aspects of the demonstrated tasks without specifically tailored primitive behaviors. We validate our approach in a simulated environment with a Pioneer 3DX mobile robot. We demonstrate the advantages of our learning approach through comparison with manually coded controllers and sequential learning