Scholarlink

<p>Physics-based neural networks are widely used in controller design for dynamic systems, as they can compensate for unknown disturbances while remaining consistent with the underlying mathematical model. The paper proposes a physics-informed neural network (PINN)-based event-triggered model predictive controller (ET-MPC) for trajectory planning in robot manipulators. A deep Lagrangian network (DeLaN) used as the PINN model is trained to estimate the dynamics of the manipulator by approximating the dynamics parameter matrices and then computing the inverse or forward relation. A two-stage training method is proposed based on the defined data and physics loss function to improve convergence, accounting for measurement noise and modelling uncertainties. The manipulator trajectory is predicted by optimizing an error-based quadratic function, subject to the imposed state, control, and collision-avoidance constraints, within the MPC framework. The computational load and task execution time are reduced by restricting the optimization based on two triggering conditions. By using the PINN-based system dynamics approximation and event-based optimization strategy, the proposed MPC controller provides improved trajectory tracking capability with reduced optimization and task time. The proposed design is practically validated using a 5-DOF ViperX300 manipulator, and the results show improved tracking efficiency of 29.5% with a 74.6% reduction in computation time and 26.5% reduction in task execution time compared to conventional analytical model-based MPC.</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

Event-triggered MPC design for manipulator trajectory planning using physics-informed neural networks

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • K Dileep,
  • S J Mija,
  • N K Arun

摘要

Physics-based neural networks are widely used in controller design for dynamic systems, as they can compensate for unknown disturbances while remaining consistent with the underlying mathematical model. The paper proposes a physics-informed neural network (PINN)-based event-triggered model predictive controller (ET-MPC) for trajectory planning in robot manipulators. A deep Lagrangian network (DeLaN) used as the PINN model is trained to estimate the dynamics of the manipulator by approximating the dynamics parameter matrices and then computing the inverse or forward relation. A two-stage training method is proposed based on the defined data and physics loss function to improve convergence, accounting for measurement noise and modelling uncertainties. The manipulator trajectory is predicted by optimizing an error-based quadratic function, subject to the imposed state, control, and collision-avoidance constraints, within the MPC framework. The computational load and task execution time are reduced by restricting the optimization based on two triggering conditions. By using the PINN-based system dynamics approximation and event-based optimization strategy, the proposed MPC controller provides improved trajectory tracking capability with reduced optimization and task time. The proposed design is practically validated using a 5-DOF ViperX300 manipulator, and the results show improved tracking efficiency of 29.5% with a 74.6% reduction in computation time and 26.5% reduction in task execution time compared to conventional analytical model-based MPC.