基于强化学习算法的四足爬行机器人运动控制  

Locomotion Control of Quadrupedal Sprawling Robots Based on Reinforcement Learning Algorithm

作　　者：曾一帆 姬宇 吴暄 孙涛 ZENG Yifan;JI Yu;WU Xuan;SUN Tao(China Nanhu Academy of Electronics and Information Technology,Jiaxing 314002,China)

机构地区：[1]中国电子科技南湖研究院,浙江嘉兴314002

出　　处：《智能物联技术》2025年第1期133-139,共7页Technology of Io T& AI

基　　金：国防基础科研计划(JCKY2023210C002)。

摘　　要：强化学习算法在四足爬行机器人运动控制中的应用较少,机器人关节构型对训练控制策略的影响尚未可知。对此,提出基于强化学习算法的四足爬行机器人运动控制方法,研究不同关节初始角度对运动控制策略收敛特性的影响。实验结果表明:机器人关节初始角度接近其极限值时,运动控制策略的收敛速度慢,收敛效果较差。以适当的关节初始角度训练的最优控制策略,可实现四足爬行机器人在草地、石子地和硬质地面上的自适应运动。The application of reinforcement learning algorithms in locomotion control for quadrupedal sprawling robots is relatively scarce,and the impact of robot joint configurations on the training of control policies remains unknown.This paper proposes a motion control method for quadruped sprawling robots based on reinforcement learning algorithms and research the influence of different initial joint angles on the convergence performance of locomotion control policies for quadrupedal sprawling robots.Experimental results indicate that when the robot's joint initial angles approximate their limit values,the convergence effect of the control policies is poorer.By implementing the control policy,which has been trained with appropriate joint initial angles,onto a robot prototype,adaptive locomotion of the quadrupedal sprawling robot on grass,gravel,and hard ground is successfully achieved.

关 键 词：近端策略优化算法 四足爬行机器人 运动控制 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]