Encoding Robot Behavior as Sensory-Based Adaptation of Learned Skillful Trajectories

Abstract

The imitation learning paradigm is a systematic approach for encoding intelligent behaviors into robotic systems. While a model representation of the ideal task behavior can be learned by processing a set of human demonstrations, learning a modeling representation that can generalize the desired behavior to perform well in dynamic environments with human collaborators is an open challenge. To address this problem, we encode intelligent robot behavior as a combination of a popular learned baseline control policy (Gaussian Mixture Model, GMM) with reactive control policies that activate based on triggers from online sensory information during task execution. Two contributions encapsulate the approach: an iterative algorithm to combine the learned and reactive policies and examples for mapping sensory information into desired robot reactive behaviors. The proposed approach was implemented on a bi-manual surgical robot and evaluated on how well the combined control policy balanced the behavioral constraints imposed during a collision avoidance and compliance tasks. Successful dynamic collision avoidance results and compliance responses that reduce environmental forces on the manipulator support the use of this paradigm for designing intelligent robot behaviors which can complement learned models to program complex robot behaviors that can balance task performance in scenarios with human collaborators.