基于近端策略优化算法及视觉感知的机械臂导纳控制研究

Research on Admittance Control of Manipulator Based on Proximal Policy Optimization Algorithm and Visual Perception

摘要: 现代机械臂交互任务中，由于环境的复杂性和不确定性，精确的物体表面建模常常难以实现。因此，如何在不依赖精确模型的情况下，提高机械臂与环境交互时的适应性和稳定性成为机械臂与环境交互任务的研究重点之一。本文针对机械臂交互任务，旨在实现视觉引导下的精细力控。研究工作首先基于MuJoCo (Multi-Joint Dynamics with Contact)物理引擎搭建了机械臂交互仿真环境，并创新性的融合了基于位置的视觉伺服(Position-Based Visual Servo, PBVS)控制和导纳控制。通过深度强化学习(Deep Reinforcement Learning, DRL)中的近端策略优化(Proximal Policy Optimization, PPO)算法，有效整合了视觉信息和力信息，从而提出了一种结合了视觉感知的导纳控制策略。通过对比实验验证，结合视觉感知的导纳控制相较于视觉伺服控制，力控整体性能提升68.75%；相较于经典的导纳控制，峰值力控制精度提高15%。实验结果表明，结合视觉感知的导纳控制在平面和不规则凹面环境中均表现出色：不仅能精确执行视觉引导下的力控任务，还能在多样化的接触面上保持稳定的交互力并迅速适应环境变化。在精密装配、医疗辅助和服务机械臂等领域，能够提高机械臂在复杂、不确定环境中的适应性和稳定性，从而推动智能机械臂自主操作的进一步发展。

Abstract: In modern manipulator interaction tasks, due to the complexity and uncertainty of the environment, accurate object surface modeling is often difficult to achieve. Therefore, improving the adaptability and stability of the interaction between the manipulator and the environment has become one of the research focuses of the interaction task. Aiming at the interactive task of the manipulator, this paper aims to realize the fine force control under visual guidance. Therefore, based on the MuJoCo (Multi-Joint Dynamics with Contact) physics engine, we built an interactive simulation environment for the manipulator, and innovatively integrated the position-based visual servo control and admittance control. Through the Proximal Policy Optimization (PPO) algorithm in Deep Reinforcement Learning (DRL), the visual information and force information are effectively integrated, and an admittance control strategy combining visual perception is proposed. Through comparative experiments, the admittance control combined with visual perception is compared with visual servo control, and the overall performance of force control is improved by 68.75%. Compared with the classical admittance control, the peak force control accuracy is improved by 15%. The experimental results showed that the admittance control combined with visual perception performs well in both flat and irregular concave environments: it can not only accurately perform visual-guided force control tasks, but also maintain stable interaction forces on a variety of contact surfaces and quickly adapt to environmental changes. In the fields of precision assembly, medical assistance and service manipulator, it can improve the adaptability and stability of manipulator in complex and uncertain environments, thus promoting the further development of autonomous operation of intelligent manipulator.