基于弹性分析与函数回归模型的示教轨迹学习与泛化方法  

Learning and generalizing demonstration trajectories based on elastic analysis and the functional regression model

作　　者：肖宏远 刘成菊[1] 陈启军[1] Hongyuan XIAO;Chengju LIU;Qijun CHEN(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804

出　　处：《中国科学：信息科学》2022年第10期1822-1836,共15页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61733013,62173248,62073245)资助项目。

摘　　要：示教学习是解决仿人机器人动态轨迹规划问题的高效方法.示教轨迹作为一种函数型数据,天然地具有耦合的时空特征,需要在函数型数据分析的理论框架下进行研究.传统的示教学习成果以“预处理”的形式解决上述耦合问题,不能充分利用示教轨迹的函数型数据本质.为了充分利用这一本质,本文采用弹性分析方法,将示教轨迹视作函数型数据,解决它们的时–空特征解耦问题.基于解耦后的示教轨迹,本文建立函数型回归模型并将其应用于动态轨迹的生成,实现示教轨迹在不同任务参数下的泛化.在实验中,JACO-2机械臂学习形态各异的技能并在不同的任务参数下完成任务,验证了提出方法的有效性.本文在统一的数学框架下,基于函数型数据分析的数学方法,解决一类仿人机器人的示教轨迹学习与泛化问题,具有较好的理论统一性与可实践性.Learning from demonstration(LfD) has been confirmed to be an effective method for solving humanoid trajectory-planning problems under dynamic scenarios. Demonstrations, as functional data, can have coupled spatial and temporal features, which need to be treated carefully under the framework of functional data analysis. Previous studies in LfD use a separate pre-processing step to decouple spatial and temporal features. To fully utilize the intrinsic functional feature of demonstrations, an elastic method is used to extract their spatial and temporal features. Based on the decoupled spatial feature, a functional regression model is built to generate dynamic trajectories under varying task parameters. Experiments on a JACO-2 manipulator are conducted by performing different skills and fulfilling related tasks dynamically with new task-parameters. Results demonstrate that a category of humanoid LfD problems can be solved with a theoretically unified and applicable method.

关 键 词：示教学习 轨迹生成 函数型数据 弹性分析 函数回归模型 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]