| 李逃昌,李健璋,侯利民,金海波.多智能体深度强化学习优化的机器人导纳控制[J].电子测量与仪器学报,2025,39(5):134-143 |
| 多智能体深度强化学习优化的机器人导纳控制 |
| Robot admittance control optimized by multi-agent deep reinforcement learning |
| |
| DOI： |
| 中文关键词: 机器人 深度强化学习 导纳控制 |
| 英文关键词:robot deep reinforcement learning admittance control |
| 基金项目:国家自然科学基金（52177047，62173171）、葫芦岛科技计划项目（2024JH(2)2/05b）资助 |
|
|
| Author | Institution |
| Li Taochang | Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao 125105, China |
| Li Jianzhang | Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao 125105, China |
| Hou Limin | Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao 125105, China |
| Jin Haibo | Faculty of Software, Liaoning Technical University, Huludao 125105, China |
|
| 摘要点击次数: 1125 |
| 全文下载次数: 761 |
| 中文摘要: |
| 针对固定参数主动柔顺控制受机器人内部参数不确定等建模误差影响导致轨迹精度不高的问题，提出一种基于多智能体深度确定性策略梯度（MA-DDPG）的机器人自适应导纳控制方法。首先，基于机器人模型建立导纳控制器。其次，将深度确定性策略梯度（DDPG）算法与导纳控制相结合，设计了一种由DDPG智能体直接输出导纳参数的自适应导纳控制器。针对其收敛速度慢和控制效果不好的问题，在自适应导纳控制算法中引入多智能体思想，将每一个导纳控制参数作为一个智能体的输出，采用集中式训练分布式执行架构的MA-DDPG算法对导纳控制器参数进行协同优化。最后，通过对比深度强化学习仿真训练效果以及自适应导纳控制在期望轨迹上的受力实验效果，验证了所提方法的可行性与有效性。实验数据表明，与其他深度强化学习算法的自适应导纳控制相比，所提方法的仿真训练收敛速度提高了65.88%，轨迹精度提高了63.35%。 |
| 英文摘要: |
| The paper proposes a robot adaptive admittance control method based on the multi-agent deep deterministic policy gradient (MA-DDPG) to address the issue of low trajectory accuracy in fixed-parameter active compliance control caused by modeling errors, such as uncertainty in robot internal parameters. Firstly, an admittance controller is established based on the robot model. Secondly, by integrating the DDPG algorithm with the admittance control framework, an adaptive admittance controller is developed, wherein the DDPG-based agent dynamically generates optimal admittance parameters. To address issues of slow convergence and unsatisfactory control performance, the concept of multiple agents is introduced into the adaptive admittance control algorithm, with each agent responsible for optimizing an individual admittance control parameter. The MA-DDPG algorithm, based on a centralized training and distributed execution architecture, is employed to optimize the admittance controller parameters. Finally, the feasibility and effectiveness of the proposed method are validated through a comparative analysis between the impact of deep reinforcement learning simulation training and the experimental outcomes of adaptive admittance control on the anticipated trajectory. The experimental data demonstrate that in comparison with adaptive admittance control based on alternative deep reinforcement learning algorithms, the proposed method exhibits a 65.88% improvement in convergence speed and a 63.35% enhancement in trajectory accuracy. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
