Scholarlink

Precise and robust trajectory tracking for quadrotors is challenged by complex dynamics and external disturbances. While model-based controllers are vulnerable to inaccuracies, standard Deep Reinforcement Learning (DRL) often yields non-smooth actions and poor robustness. This paper proposes Predictive Error-Augmented Reinforcement Learning (PEARL) to address these dual challenges. PEARL integrates two key mechanisms: a predictive error module to actively compensate for model-reality mismatch and a Lipschitz-constrained policy to architecturally enforce control smoothness. Experiments in a high-fidelity simulator show PEARL achieves tracking precision comparable to an oracle Model Predictive Controll (MPC). Furthermore, it significantly outperforms a standard DRL baseline in disturbance rejection and control smoothness. PEARL thus offers a principled path toward reliable and deployable DRL controllers for safety-critical quadrotors applications.

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

Predictive Error-Augmented Reinforcement Learning for Robust and Smooth Quadrotor Trajectory Tracking

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bo Wang,
  • Mengyun Wang,
  • Huangzhi Yu,
  • Yifeng Niu

摘要

Precise and robust trajectory tracking for quadrotors is challenged by complex dynamics and external disturbances. While model-based controllers are vulnerable to inaccuracies, standard Deep Reinforcement Learning (DRL) often yields non-smooth actions and poor robustness. This paper proposes Predictive Error-Augmented Reinforcement Learning (PEARL) to address these dual challenges. PEARL integrates two key mechanisms: a predictive error module to actively compensate for model-reality mismatch and a Lipschitz-constrained policy to architecturally enforce control smoothness. Experiments in a high-fidelity simulator show PEARL achieves tracking precision comparable to an oracle Model Predictive Controll (MPC). Furthermore, it significantly outperforms a standard DRL baseline in disturbance rejection and control smoothness. PEARL thus offers a principled path toward reliable and deployable DRL controllers for safety-critical quadrotors applications.