Scholarlink

<p>Robust and adaptive control of robotic manipulators is critical for managing uncertainties and disturbances. Traditional control methods often lack adaptability to varying dynamic conditions, necessitating advanced approaches that ensure stability and performance. This paper presents two reinforcement learning (RL) schemes designed to enhance the robustness and adaptability of a robotic manipulation system using both model-based and model-free control approaches. A nonlinear control design integrates a proportional-derivative controller with a finite-time synergetic controller in the model-based controller, effectively addressing diverse object models and uncertainties in robotic dynamics. The first RL scheme (RL1) tunes the control parameters of this nonlinear controller, enabling real-time adaptation to system uncertainties and ensuring stability through a Lyapunov-based analysis. For the model-free controller, a multi-layer neural network is developed using a model-driven approach to predict actuated torque, compensating for the nominal controller. The second RL scheme (RL2) optimizes the neural network for torque prediction, enhancing control performance during robot operations. Experimental validation on a parallel Delta robot subjected to dynamic trajectories and external disturbances demonstrates the effectiveness and flexibility of the proposed methods during episodic operations. With the resolution of encoders as 0.0064 rad per pulse, RL1 achieved the best RMS trajectory errors of 0.0066 rad and a maximum error of 0.0175 rad after only 60 training episodes. RL2 successfully optimized torque compensation with the most optimized RMS error of 0.0069 rad and a maximum error of 0.0160 rad, requiring 150 training episodes for stable convergence. Both RL algorithms are optimized for real-time implementation and safety, showcasing their feasibility in practical robotic applications.</p>

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

Online Reinforcement Learning Based Real-time Robust Adaptive Control Design for Robot Manipulators

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • Phu-Cuong Pham,
  • Yong-Lin Kuo

摘要

Robust and adaptive control of robotic manipulators is critical for managing uncertainties and disturbances. Traditional control methods often lack adaptability to varying dynamic conditions, necessitating advanced approaches that ensure stability and performance. This paper presents two reinforcement learning (RL) schemes designed to enhance the robustness and adaptability of a robotic manipulation system using both model-based and model-free control approaches. A nonlinear control design integrates a proportional-derivative controller with a finite-time synergetic controller in the model-based controller, effectively addressing diverse object models and uncertainties in robotic dynamics. The first RL scheme (RL1) tunes the control parameters of this nonlinear controller, enabling real-time adaptation to system uncertainties and ensuring stability through a Lyapunov-based analysis. For the model-free controller, a multi-layer neural network is developed using a model-driven approach to predict actuated torque, compensating for the nominal controller. The second RL scheme (RL2) optimizes the neural network for torque prediction, enhancing control performance during robot operations. Experimental validation on a parallel Delta robot subjected to dynamic trajectories and external disturbances demonstrates the effectiveness and flexibility of the proposed methods during episodic operations. With the resolution of encoders as 0.0064 rad per pulse, RL1 achieved the best RMS trajectory errors of 0.0066 rad and a maximum error of 0.0175 rad after only 60 training episodes. RL2 successfully optimized torque compensation with the most optimized RMS error of 0.0069 rad and a maximum error of 0.0160 rad, requiring 150 training episodes for stable convergence. Both RL algorithms are optimized for real-time implementation and safety, showcasing their feasibility in practical robotic applications.