Scholarlink

<p>This paper proposes an adaptive Neural Network–based Nonlinear Model Predictive Control (NN–NMPC) framework to enhance the path tracking accuracy and computational efficiency of a four-wheeled mobile robot in dynamic environments. The NMPC gain matrices are tuned online using a compact 4 × 5 multilayer perceptron, trained on 18,360 input–output samples with up to 29-step future reference data and noise augmentation for robustness. Experimental evaluations over velocities from 0.5&#xa0;m/s to 1.5&#xa0;m/<b>s</b> demonstrate a 25% average tracking accuracy improvement<b>,</b> 30% reduction in computation time, and final positional error decrease from 0.10&#xa0;m to 0.07&#xa0;m compared to a fixed-gain NMPC. On complex sigma-shaped paths, the proposed method halves solution time (from 8.34&#xa0;s to 5.61&#xa0;s) without performance loss. Speed sensitivity tests show RMSE rising from 0.095&#xa0;m to 0.132&#xa0;m and noise robustness declining from 0.88 to 0.80 as speed increases, while computation time remains constant (≈ 1.5&#xa0;s). In obstacle-avoidance scenarios, the NN–NMPC with probabilistic collision-avoidance constraints maintain over 95% collision-free confidence and reduces peak deviation by up to 3 × compared to baseline NMPC under noisy perception. Real-world implementation using a Pixy2 vision system (5.5&#xa0;mm static accuracy at 2&#xa0;m) and ARM–U2D2 processing confirms the controller’s real-time feasibility and robustness. These results establish NN–NMPC as a strong candidate for intelligent, high-speed navigation in wheeled robotics.</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

Neural network-based intelligent path tracking for nonlinear predictive control in wheeled robots

  • 期刊论文

  • 发表日期：

    2025年11月6日

查看全文
  • Mostafa Jalalnezhad

摘要

This paper proposes an adaptive Neural Network–based Nonlinear Model Predictive Control (NN–NMPC) framework to enhance the path tracking accuracy and computational efficiency of a four-wheeled mobile robot in dynamic environments. The NMPC gain matrices are tuned online using a compact 4 × 5 multilayer perceptron, trained on 18,360 input–output samples with up to 29-step future reference data and noise augmentation for robustness. Experimental evaluations over velocities from 0.5 m/s to 1.5 m/s demonstrate a 25% average tracking accuracy improvement, 30% reduction in computation time, and final positional error decrease from 0.10 m to 0.07 m compared to a fixed-gain NMPC. On complex sigma-shaped paths, the proposed method halves solution time (from 8.34 s to 5.61 s) without performance loss. Speed sensitivity tests show RMSE rising from 0.095 m to 0.132 m and noise robustness declining from 0.88 to 0.80 as speed increases, while computation time remains constant (≈ 1.5 s). In obstacle-avoidance scenarios, the NN–NMPC with probabilistic collision-avoidance constraints maintain over 95% collision-free confidence and reduces peak deviation by up to 3 × compared to baseline NMPC under noisy perception. Real-world implementation using a Pixy2 vision system (5.5 mm static accuracy at 2 m) and ARM–U2D2 processing confirms the controller’s real-time feasibility and robustness. These results establish NN–NMPC as a strong candidate for intelligent, high-speed navigation in wheeled robotics.