Policy Adaptation through Tactile Correction

摘要

Abstract. Behavior adaptation based on execution experience can be a practical tool to increase the robustness of a robot behavior learned from demonstration. While demonstration learning is a pow-erful technique for the development of robot behaviors, in general de-velopment remains a challenge. This work presents an approach for policy improvement through a tactile interface located on the body of the robot. We introduce the Tactile Policy Correction (TPC) al-gorithm, that employs tactile feedback for the adaptation of a policy learned from demonstration. We provide an initial validation of re-finement under the TPC algorithm on humanoid robot performing a grasp positioning task, and policy performance is found to improve with tactile corrections. We additionally show different modalities, namely teleoperation and tactile corrections, to provide information about allowable variability in the target behavior in different areas of the state space. 1