基于改进ProMPs的机器人姿态轨迹模仿学习  

Robot Orientation Trajectory Imitation Learning Based on Improved ProMPs

作　　者：伍家俊 黎奕辉 陈燊豪 WU Jiajun;LI Yihui;CHEN Shenhao(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广州510006

出　　处：《组合机床与自动化加工技术》2024年第10期46-49,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(62203126);广东省重点研发计划项目(2020B090928002)。

摘　　要：针对概率运动基元(ProMPs)在学习机器人姿态轨迹时忽略数据背后的非线性约束导致姿态轨迹推断精度低的问题,提出了一种黎曼流形上的改进ProMPs方法。该方法采用单位四元数作为机器人姿态表达方式。首先,将ProMPs的学习过程与黎曼流形上的极大似然估计与对数映射相结合实现示教姿态轨迹的概率建模;然后,利用黎曼流形上的指数映射对ProMPs的推断轨迹还原为机器人姿态轨迹;最后,在UR5机器人进行了实验评估。结果表明,该算法相较于原始ProMPs在姿态轨迹推断精度与轨迹平滑度方面分别提升56%与35%,并能使机器人更快地到达目标姿态。Aiming at the problem that the Probabilistic Movement Primitives(ProMPs)ignore the nonlinear 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;Subsequently,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 inferred 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 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]