Abstract: The quality of regression path in collaborative robot teaching learning is limited by the distribution of teaching trajectories,and improper preprocessing may lead to overfitting of regression path or loss of task features.A path optimization strategy for segmented adaptive Gaussian noise divergence was proposed to address this issue.Considering the distribution differences of teaching trajectory various parts,a trajectory segmentation method based on feature driven clustering was proposed to divide the teaching trajectory into multiple task feature segments.To improve the distribution of the teaching trajectory and enhance the quality of the regression path,a optimization model of multi-dimensional Gaussian noise parameter was built to select appropriate divergence parameters for each feature segment.The proposed method was experimentally validated with the teaching trajectories collected by the UR5 robot,and the similarity,the feature preservation and the smoothness of the regression paths were evaluated.The results showed that the optimized regression path maintained a high degree of similarity with the original teaching trajectories as a whole.At the same time,while retaining the task features of the original trajectories,the twists and redundancies caused by adverse features were reduced,and the smoothness was further improved.

Key words: collaborative robot, learning from demonstration, trajectory segmentation, Gaussian noise divergence, path optimization

摘要： 协作机器人示教学习的回归路径质量受限于示教轨迹的分布,不当的预处理可能导致回归路径过拟合或任务特征丢失。针对此问题,提出了一种分段自适应高斯噪声发散的路径优化策略。首先,考虑到示教轨迹各部分的分布差异,提出一种基于特征驱动聚类的轨迹段分割方法将示教轨迹分割成多个任务特征段。其次,构建了多维高斯噪声参数优化模型,为每个特征段选取适当的发散参数,以改善示教轨迹的分布,进而提高回归路径的质量。最后,通过UR5机器人采集的示教轨迹对所提出的方法进行了实验验证,评估了回归路径的相似性、特征保留和平滑性。结果表明,经过优化后得到的回归路径整体上与原始示教轨迹保持了较高的相似性,同时在保留原始轨迹的任务特征的基础上,减少了由于不利特征影响而出现的曲折和冗余,平滑性得到进一步提高。

关键词: 协作机器人, 示教学习, 轨迹分割, 高斯噪声发散, 路径优化