Scholarlink

To address chatter vibration in robotic thin-walled part grinding, this paper proposes a variable impedance control (VIC) method fusing deep deterministic policy gradient (DDPG) reinforcement learning with an imitation mechanism (IM). The method establishes dynamic mapping between vibration states and impedance parameters, which forms an integrated sensing-control scheme. Within the reinforcement learning (RL) framework, a multimodal state space is constructed by combining time-frequency domain features, including wavelet packet energy and force error. An imitation mechanism is introduced to generate state-action sample sets compliant with vibration dynamics for network pre-training, which effectively resolves the safety risks and training inefficiency issues encountered by reinforcement learning in real-world environments. Experimental results demonstrate that, compared with traditional impedance control (TIC), the proposed method suppresses characteristic chatter frequency amplitudes by over 34% and reduces force control errors by 50%. These results confirm its feasibility and effectiveness. This data-driven approach achieves closed-loop chatter detection and suppression without requiring precise system dynamics modeling, providing a novel solution for precision robotic machining in flexible manufacturing scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robotic Grinding of Thin-Walled Parts: Reinforcement Learning-Based Chatter Suppression Method

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Fuyong Zhang,
  • Zhihao Xu,
  • Yuming Li,
  • Haifei Zhu,
  • Zhaoyang Liao,
  • Hongmin Wu,
  • Xubin Lin,
  • Xuefeng Zhou

摘要

To address chatter vibration in robotic thin-walled part grinding, this paper proposes a variable impedance control (VIC) method fusing deep deterministic policy gradient (DDPG) reinforcement learning with an imitation mechanism (IM). The method establishes dynamic mapping between vibration states and impedance parameters, which forms an integrated sensing-control scheme. Within the reinforcement learning (RL) framework, a multimodal state space is constructed by combining time-frequency domain features, including wavelet packet energy and force error. An imitation mechanism is introduced to generate state-action sample sets compliant with vibration dynamics for network pre-training, which effectively resolves the safety risks and training inefficiency issues encountered by reinforcement learning in real-world environments. Experimental results demonstrate that, compared with traditional impedance control (TIC), the proposed method suppresses characteristic chatter frequency amplitudes by over 34% and reduces force control errors by 50%. These results confirm its feasibility and effectiveness. This data-driven approach achieves closed-loop chatter detection and suppression without requiring precise system dynamics modeling, providing a novel solution for precision robotic machining in flexible manufacturing scenarios.