基于深度强化学习的四足机器人后空翻动作生成方法  

A Method for Generating Quadruped Backflip Motion Based on Deep Reinforcement Learning

作　　者：李岸荞 王志成 古勇[1,2] 吴俊[1,2] 朱秋国[1,2] LI An-qiao;WANG Zhi-cheng;GU Yong;WU Jun;ZHU Qiu-guo(The Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou 310027, China;State Key Laboratory of Industrial Control Technology, Hangzhou 310027, China)

机构地区：[1]浙江大学智能系统与控制研究所,杭州310027 [2]浙江大学工业控制技术国家重点实验室,杭州310027

出　　处：《导航定位与授时》2021年第6期35-42,共8页Navigation Positioning and Timing

基　　金：国家重点研发计划项目(2018YFB1305900)。

摘　　要：四足机器人灵巧运动技能的生成一直受到机器人研究者们的广泛关注,其中空中翻滚运动既能展现四足机器人运动的灵活性又具有一定的实用价值。近年来,深度强化学习方法为四足机器人的灵巧运动提供了新的实现思路,利用该方法得到的闭环神经网络控制器具有适应性强、稳定性高等特点。本文在绝影Lite机器人上使用基于模仿专家经验的深度强化学习方法,实现了仿真环境中四足机器人的后空翻动作学习,并进一步证明了设计的后空翻闭环神经网络控制器相比于开环传统位置控制器具有适应性更高的特点。The dexterous motion of quadruped robot has been widely concerned by robot researchers.Its tumbling motion in the air can not only show the flexibility of the leg-foot robot system itself,but also has certain practical value.In recent years,deep reinforcement learning method has provided new ideas for the dexterous movement of quadruped robot.The neural network controller obtained by deep reinforcement learning method has the characteristics of strong generalization and high stability.In this paper,we use the deep reinforcement learning method based on imitating the expert experience on the Jueying Lite robot platform,and realize the backflip motion learning of this quadruped robot in the simulation environment.We further prove that the obtained backflip motion controller has the characteristics of high adaptability.

关 键 词：四足机器人 后空翻 深度强化学习 神经网络 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]