基于RNN的机械臂任务模仿系统  

Robot Arm Task Imitation System Based on RNN

作　　者：于建均 吴鹏申 左国玉 阮晓钢 张远 YU Jianjun;WU Pengshen;ZUO Guoyu;RUAN Xiaogang;ZHANG Yuan(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]计算智能与智能系统北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2018年第11期1401-1408,共8页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61375086);高等学校博士学科点专项科研基金资助项目(20101103110007);北京工业大学智能制造领域大科研推进计划资助项目(JZ041001201702)

摘　　要：为了简化机械臂复杂的运动规划问题,且使机械臂具有适应新任务的泛化能力,研究并实现了一种基于循环神经网络(recurrent neural network,RNN)的机械臂任务模仿系统.首先,由示教者进行原始任务示教并采集示教数据;其次,通过构建RNN对原始示教数据进行训练,得到机械臂对示教任务模仿的控制策略;然后,当任务发生变化时,观察新任务的运动并采集运动信息;最后,通过基于RNN的控制策略对新任务运动信息进行泛化输出,得到机械臂模仿新任务的控制信息,进而完成模仿.物理对象实验结果表明,系统具有简单高效的策略获取能力以及良好的泛化能力,使机械臂不仅能够模仿原始示教任务,而且可以通过泛化实现对新任务的模仿.To simplify the complex motion planning problem of robot arm and make the robot arm have the generalization ability to adapt to the new task,a robot arm task imitation system based on recurrent neural network（ RNN） was researched and implemented in this paper. First,the original task was taught by the teacher,and the teaching data was collected. Second,by constructing RNN to train the original teaching data,the control strategy of the robot arm imitation was obtained. Then,when the task changes,the movement of the new task was observed and the movement information was collected. Finally,the motion information was generalized via the robot arm control strategy based on RNN,and the control information of robot arm of imitating the new task was obtained,completing the imitation. Experimental results show that the method can obtain strategy simply and efficiently and has preferable generalization ability,which make the robot arm not only can imitate the original task,but also can imitate the new task when the task changes.

关 键 词：控制策略 机械臂 运动规划 循环神经网络(RNN) 泛化 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]