Various Robot Motor Skills Learning with PI2-GMR

Abstract

Learning from demonstration has been applied successfully in acquiring similar motor skills for robot. However, how to accomplish different tasks with no explicit demonstration is still a challenging issue. In this paper, we propose a novel robot skills learning method consisted of Dynamical Movement Primitives with mixture Gaussian Model Regression(DMPS-GMR) and Policy Improvement with Path Integrals (PI2). The DMPS-GMR make the robot have the ability of learning fundamental task from the rough demonstration, and then Policy Improvement with Path Integrals based on GMR (PI2-GMR) endow robot the optimal/suboptimal solution for dissimilar task from the imitated state gain from DMPS-GMR. Experimental results demonstrate that the proposed approach can make robot acquisition skill more accurately.