Scholarlink

<p>This study proposes a novel self-attention network (SAN)-based auto-tuning framework for model predictive control (MPC) in mobile robot systems. Compared with conventional manual parameter tuning methods, the proposed framework leverages machine learning to dynamically adapt both MPC weight matrices and prediction horizons in real-time, thereby enhancing both system adaptability and computational efficiency. Firstly, a linearized model is derived to characterize the robot’s state-space behavior, enabling explicit analysis of the coupling between MPC parameters and system dynamics. Secondly, a SAN module is designed to dynamically tune MPC weight matrices in real-time, while an independent prediction horizon reduction strategy gradually decreases horizon length during operation. Thirdly, an event-triggered mechanism coordinates weight updates and prediction horizon resets based on disturbance monitoring. Finally, the effectiveness of the proposed framework is validated through path tracking experiments on a real mobile robot across three diverse trajectories. Results show that for each of the three trajectories, the proposed Auto-tuning MPC reduced the average evaluation value by 36.7%, 64.1%, and 77.4% compared to conventional MPC and achieved 67.6%, 48.9%, and 77.1% lower computation time compared to BP-tuning 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

Self-attention network-based auto-tuning MPC for mobile robot systems

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Yuesheng Liu,
  • Zhongxian Xu,
  • Ning He,
  • Lile He,
  • Fuan Cheng

摘要

This study proposes a novel self-attention network (SAN)-based auto-tuning framework for model predictive control (MPC) in mobile robot systems. Compared with conventional manual parameter tuning methods, the proposed framework leverages machine learning to dynamically adapt both MPC weight matrices and prediction horizons in real-time, thereby enhancing both system adaptability and computational efficiency. Firstly, a linearized model is derived to characterize the robot’s state-space behavior, enabling explicit analysis of the coupling between MPC parameters and system dynamics. Secondly, a SAN module is designed to dynamically tune MPC weight matrices in real-time, while an independent prediction horizon reduction strategy gradually decreases horizon length during operation. Thirdly, an event-triggered mechanism coordinates weight updates and prediction horizon resets based on disturbance monitoring. Finally, the effectiveness of the proposed framework is validated through path tracking experiments on a real mobile robot across three diverse trajectories. Results show that for each of the three trajectories, the proposed Auto-tuning MPC reduced the average evaluation value by 36.7%, 64.1%, and 77.4% compared to conventional MPC and achieved 67.6%, 48.9%, and 77.1% lower computation time compared to BP-tuning MPC.