Dynamic Movement Primitives for Movement Generation Using GMM-GMR Analytical Method

Abstract

Motion generalization is an effective way for robot leaner to learn from demonstration, especially they are set within a novel situation. However, as for learned skills, to generate humanoid and natural behaviour for robot is the key challenge in robot skill learning. In this paper, we proposed a method using the statistical method Gaussian mixture model and Gaussian mixture regression (GMM-GMR) to analyze the data from human demonstration. For accurate learning, the raw data is pretreated by dynamic time warping (DTW). Dynamic movement primitives (DMP) aim to generate a human-like motion to a new goal, employing the data processed by GMM-GMR. Including induction, summarizing demonstration data and generalizing skill, the results, in comparison with Average method pretreating data, show that our method can achieve task-specific generalization with more smooth and human-like trajectory.