Scholarlink

With the upgrading of flexible manufacturing requirements in Industry 4.0, traditional robotic arm control methods face challenges such as heavy parameter tuning workload and poor adaptability to unmodeled dynamics (e.g., joint friction, sudden load changes). Existing deep reinforcement learning (DRL) methods, while promising, suffer from inefficient integration of path planning and real-time control, slow learning of inverse kinematics, and simplistic fusion of multimodal state information (position, velocity, acceleration), limiting their performance in dynamic tasks like peg-in-hole (PiH) assembly. To address these issues, this study proposes an impedance control-coupled DRL framework that integrates real-time control with path planning, and designs a multimodal attention feature fusion network to enhance the utilization of multimodal state information. Built on the Soft Actor-Critic (SAC) algorithm and combined with curriculum learning to accelerate training, the framework is validated in the PyBullet simulation environment using the FAIRINO FR5 6-axis robotic arm. Experimental results show that compared to a control group with a conventional neural network structure, the proposed attention-equipped DRL model reaches 0.1 mm positioning accuracy at 5 million training steps, far exceeding the 0.6 mm limit of the control group. It provides a new paradigm for integrating DRL with impedance control, verifying the effectiveness of multimodal attention mechanisms in improving control precision and adaptability.

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

Multimodal Attention-Based Impedance Reinforcement Learning for Robotic Peg-In-Hole Assembly

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Taoyuan Zhang,
  • Haoran Ke,
  • Guisheng Fang,
  • Weijia Wang,
  • Chuanyuan Zhou,
  • Yachong Zhou,
  • Xiaoxiao Wei,
  • Jing Nie

摘要

With the upgrading of flexible manufacturing requirements in Industry 4.0, traditional robotic arm control methods face challenges such as heavy parameter tuning workload and poor adaptability to unmodeled dynamics (e.g., joint friction, sudden load changes). Existing deep reinforcement learning (DRL) methods, while promising, suffer from inefficient integration of path planning and real-time control, slow learning of inverse kinematics, and simplistic fusion of multimodal state information (position, velocity, acceleration), limiting their performance in dynamic tasks like peg-in-hole (PiH) assembly. To address these issues, this study proposes an impedance control-coupled DRL framework that integrates real-time control with path planning, and designs a multimodal attention feature fusion network to enhance the utilization of multimodal state information. Built on the Soft Actor-Critic (SAC) algorithm and combined with curriculum learning to accelerate training, the framework is validated in the PyBullet simulation environment using the FAIRINO FR5 6-axis robotic arm. Experimental results show that compared to a control group with a conventional neural network structure, the proposed attention-equipped DRL model reaches 0.1 mm positioning accuracy at 5 million training steps, far exceeding the 0.6 mm limit of the control group. It provides a new paradigm for integrating DRL with impedance control, verifying the effectiveness of multimodal attention mechanisms in improving control precision and adaptability.