Aiming at the problem that the Probabilistic Movement Primitives(ProMPs)ignore the nonlin-ear constraints behind the data when learning robot orientation trajectory,an improved ProMPs algorithm on Riemannian manifolds is proposed.This method exploits unit quaternion to describe the robot orientation.Firstly,the learning process of ProMPs is combined with the logarithmic mapping and Maximum Likelihood Estimate on Riemannian manifolds to probabilistically model the demonstrated orientation trajectories;Sub-sequently,the exponential mapping on Riemannian manifolds is exploited to convert the ProMPs inferred trajectory into the robot orientation trajectory;Finally,the algorithm performance is evaluated based on the UR5 robot.Experiment results show that the improved ProMPs enhance the accuracy and smoothness of in-ferred unit quaternion trajectories by 56%and 35%,respectively,when compared to the original ProMPs.The improved ProMPs also allow the robot to reach the target orientation faster.
关键词
机器人/模仿学习/单位四元数/概率运动基元/黎曼流形
Key words
robots/imitation learning/unit quaternion/probabilistic movement primitives/riemannian mani-folds
维普
万方数据