High-Speed Trajectory Tracking Control for Quadrotors via Deep Reinforcement Learning

• Main Authors: Hua, Hean, Zhong, Hang, Zhang, Hui, Fang, Yongchun, Wang, Yaonan
• Format: Conference Proceeding
• Language: English
• Subjects: Adaptation models; cooperative learning; Deep reinforcement learning; Hardware; Heuristic algorithms; high-speed flight; Optimal control; practical application; Quadrotors; Training; Trajectory tracking
• ISSN: 2837-8601
• DOI: 10.1109/YAC63405.2024.10598678

This paper presents a learning-based high-speed trajectory tracking control strategy for quadrotors, which achieves efficient learning and strong reliability by the collaboration of deep reinforcement learning (RL) and self-tuning mechanism. Different from existing methods, the proposed strategy is designed to explore optimal control performance by taking advantage of model-based self-tuning mechanism and deep reinforcement learning. Specifically, the self-tuning guided deep RL scheme is put forward for quadrotors, with superior learning efficiency and strong adaptability. Firstly, a novel self-tuning mechanism is constructed and some auxiliary variables are introduced to enhance the tracking performance. Then, based on the model-driven self-tuning design, the deep RL is proposed to achieve model-guided learning, where the tuning actions are adopted in the evaluation process during training, aiming at removing the bad explorations by the carefully designed parallel evaluation. Finally, the convergence is analyzed based on the proposed learning framework, which indicates the efficient cooperation of exploration and self-tuning mechanism. To verify the effectiveness of the proposed controller, the guided training and hardware experiments are implemented to show efficient cooperation and satisfactory high-speed trajectory tracking control of the proposed method.