Scholarlink

<p>To achieve precise control of a system with unknown friction, this paper proposes a control method that integrates reinforcement learning (RL)-based friction compensation with linear quadratic regulation (LQR) control for cart-inverted pendulum systems. Specifically, we integrate the mathematical models of Coulomb and viscous friction into the dynamic model of the cart-inverted pendulum. Since the friction coefficients are uncertain in practice, an RL agent is employed to estimate these unknown coefficients. The soft actor-critic (SAC) algorithm is utilized to estimate the coefficients of the Coulomb and viscous friction models, as it trains the proposed RL agent in a stable and efficient manner. The RL-based friction compensation is incorporated into the LQR controller, which is designed using a nominal linear model, thereby significantly reducing the computational complexity of training the RL agent. The effectiveness of the proposed method in addressing scenarios with unknown friction is evaluated through comprehensive simulations and experiments using three distinct reward functions by comparing it with conventional methods. The proposed method achieved a root mean square error (RMSE) for the pendulum angle as low as 0.0033, whereas the best-performing conventional method exhibited an RMSE of 0.0073, representing at least a 48% reduction in experimental results.</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

Reinforcement learning-based friction compensation of an inverted pendulum on a cart

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Jaehwan Jeong,
  • Jaepil Ban

摘要

To achieve precise control of a system with unknown friction, this paper proposes a control method that integrates reinforcement learning (RL)-based friction compensation with linear quadratic regulation (LQR) control for cart-inverted pendulum systems. Specifically, we integrate the mathematical models of Coulomb and viscous friction into the dynamic model of the cart-inverted pendulum. Since the friction coefficients are uncertain in practice, an RL agent is employed to estimate these unknown coefficients. The soft actor-critic (SAC) algorithm is utilized to estimate the coefficients of the Coulomb and viscous friction models, as it trains the proposed RL agent in a stable and efficient manner. The RL-based friction compensation is incorporated into the LQR controller, which is designed using a nominal linear model, thereby significantly reducing the computational complexity of training the RL agent. The effectiveness of the proposed method in addressing scenarios with unknown friction is evaluated through comprehensive simulations and experiments using three distinct reward functions by comparing it with conventional methods. The proposed method achieved a root mean square error (RMSE) for the pendulum angle as low as 0.0033, whereas the best-performing conventional method exhibited an RMSE of 0.0073, representing at least a 48% reduction in experimental results.