摘要
针对传统机械臂控制方法对精确建模的依赖性和环境适应性不足问题,研究了基于深度强化学习的机械臂点位运动控制方法。构建了基于PyBullet的6自由度机械臂仿真平台,设计了24维状态空间和6维连续动作空间,并采用多组件奖励函数引导学习。重点对比了深度确定性策略梯度算法(Deep Deterministic Policy Gradient,DDPG)和双延迟深度确定性策略梯度算法(Twin Delayed Deep Deterministic Policy Gradient,TD3)的控制性能。实验结果表明,TD3算法通过引用三个技术能够有效抑制Q值过估计,收敛速度较DDPG提升3~5倍,成功率稳定在98.2%,显著优于DDPG的85%。研究验证了TD3算法在连续控制任务中的优越性,为深度强化学习在机器人控制领域应用提供了技术参考。
To address the dependence on precise modeling and insufficient environmental adaptability of traditional robotic arm control methods,this paper investigates point-to-point control methods based on deep reinforcement learning.A 6-DOF robotic arm simulation platform was constructed using PyBullet,with a 24-dimensional state space and 6-dimensional continuous action space designed,employing multi-component reward functions to guide learning.The control performance of DDPG and TD3 algorithms was systematically compared.Experimental results demonstrate that TD3 algorithm effectively suppresses Q-value overestimation through twin Q-networks,delayed policy updates,and target policy smoothing techniques,achieving 3-5 times faster convergence than DDPG with a stable success rate of 98.2%,significantly outperforming DDPG's 85%.This study validates the superiority of TD3 algorithm in continuous control tasks,providing technical reference for deep reinforcement learning applications in robotic control.
作者
王瑞
董枫
向明
汪伟星
Wang Rui;Dong Feng;Xiang Ming;Wang Weixin(Jingdezhen Ceramic University,College of mechanical and Electronic Engineering,Jingdezhen Jiangxi 333403,China;Yichun Hongzhou Ceramics Limited Company,Yichun Jiangxi 336023,China)
出处
《机械管理开发》
2025年第12期286-290,共5页
Mechanical Management and Development
关键词
深度强化学习
机械臂控制
TD3算法
连续控制
点位运动
deep reinforcement learning
robotic arm control
DDPG algorithm
TD3 algorithm
continuous control
point-to-point motion
