一种基于DTW-DP-GMM的工业机器人轨迹学习策略  

A Strategy for Industrial Robot Trajectory Learning Based on DTW-DP-GMM

作　　者：肖洒[1,2] 陈旭阳 叶锦华 吴海彬[1,2] Xiao Sa;Chen Xuyang;Ye Jinhua;Wu Haibin(School of Mechanical Engineering and Automation,Fuzhou University,Fuzhou 350108,China;Fujian Key Laboratory of Special Intelligent Equipment Safety Measurement and Control,Fuzhou 350108,China)

机构地区：[1]福州大学机械工程及自动化学院,福州350108 [2]福建省特种智能装备安全与测控重点实验室,福州350108

出　　处：《天津大学学报（自然科学与工程技术版）》2025年第1期68-80,共13页Journal of Tianjin University：Science and Technology

基　　金：国家重点研发计划资助项目(2018YFB1308603);福建省高校产学合作项目(2022H6016)。

摘　　要：针对机器人示教编程过程中使用高斯混合模型(GMM)规划运动轨迹时存在的高斯分布个数难以选择、复现轨迹精度较低等问题,提出了一种复合的机器人运动轨迹学习策略.该策略包含动态时间规整(DTW)算法、高斯混合模型与道格拉斯-普克(DP)算法.首先,针对示教过程中采集的多条轨迹在时间长度上存在差异的问题,采用DTW算法来统一示教轨迹在时域上的变化.其次,使用GMM算法对示教轨迹的特征进行提取,并利用高斯混合回归(GMR)算法将其重构为复现轨迹.在这个过程中采用DP算法来预估GMM算法的关键参数高斯分布的数量,与传统方法相比,能够简单直观地得到相对准确的参数值.利用DP算法对复现轨迹的数据点进行稀疏化并优化,不仅确保了机器人最终运动轨迹的精度,而且大幅减少了最终轨迹数据点的数量.最后,进行了不同形状的模拟焊接轨迹学习规划实验.结果表明:经由DTW对齐后的示教轨迹具有更加明显的运动特征,经过GMM-GMR学习输出的复现轨迹具有良好的表征结果;在使用GMM-GMR算法学习示教轨迹的过程中,采用DP算法可以有效预估高斯分布个数;经过DP算法稀疏化并优化的最终轨迹的平均位置误差均在0.500 mm以内,其最大误差可以控制在0.800 mm以内,可以满足焊接轨迹规划的精度要求,验证了该策略的有效性和优越性.The users often encounter issues such as difficulty in selecting the appropriate number of Gaussian distributions and low accuracy in reproducing trajectories when using Gaussian mixture model(GMM)to plan robot trajectories during programming by demonstration.To address these concerns,a composite strategy is proposed,which integrates dynamic time warping(DTW)algorithm,GMM and the Douglas-Peucker(DP)algorithm.First,to address the issue of varying time lengths in multiple trajectories,the DTW algorithm is used to align the variation of the demonstrated trajectories in the time domain.Second,the motion features are learned from the aligned demonstrated trajectories using GMM,which can subsequently be reconstructed into a reproduced trajectory using Gaussian mixture regression(GMR).In this process,the number of Gaussian distributions,a key parameter of GMM,is estimated by DP algorithm,which can derive a relatively precise parameter value simply and intuitively compared with the traditional method.Furthermore,the DP algorithm is employed to sparsify and optimize the data points in the reproduced trajectory,ensuring that the final trajectory maintains high precision while drastically reducing the number of data points in the final trajectory.Finally,experiments conducted on simulated welding trajectories of different shapes are carried out.The experimental results show that the demonstrated trajectories aligned by DTW exhibit more pronounced motion features,and the reproduced trajectory generated using GMM-GMR has great representation result;moreover,the DP algorithm effectively estimates the necessary number of Gaussian distributions for GMM-GMR learning.The average positional errors in final trajectories sparsified by the DP algorithm are within 0.500 mm,and the maximum errors can be controlled within 0.800 mm,meeting the precision requirements of weld-ing trajectory planning.It verifies the effectiveness and the superiority of the proposed strategy.

关 键 词：工业机器人 示教编程 高斯混合模型 道格拉斯-普克算法 动态时间规整 轨迹复现 

分 类 号：TP242.2[自动化与计算机技术—检测技术与自动化装置]