关键词: 旋翼无人机, 飞行控制, 模仿学习, 强化学习, 鲁棒控制

Abstract: In this paper, an imitation-learning-driven robust flight control method is proposed for multirotor UAVs. Prior flight data collected by a nonlinear disturbance observer are combined with deep reinforcement learning exploration to improve training efficiency and enhance robustness against external disturbances and model uncertainties. Based on the multirotor dynamic model and a hierarchical attitude--position control framework, the flight control task is formulated as a robust trajectory-tracking problem, and an observer-based control scheme is developed to estimate uncertainties online and generate demonstration data for imitation learning. A Deep Deterministic Policy Gradient (DDPG) network is then introduced to learn the robust flight control policy. The training process consists of two stages. In the first stage, the network is offline initialized using observer-generated trajectories and a hybrid loss function that accounts for both single-step and multi-step learning errors. In the second stage, the network is further optimized through exploration-driven reinforcement learning with prioritized experience replay over mixed samples from the demonstration buffer and the exploration buffer. By combining model-based prior data with online exploration data, the proposed imitation--exploration learning scheme improves both tracking accuracy and robustness. Simulation and real-flight experiments show that the proposed method achieves stable flight and effective trajectory tracking under different flight tasks and disturbance conditions. In the spiral trajectory-tracking task, the tracking errors along all three axes remain below 0.1~m. In the high-speed flight task, the maximum flight speed exceeds 2~m/s within 6~s. In the turning flight task, the measured flight speed follows the reference trend well. Real-flight results further show good agreement between the actual and desired trajectories, confirming the effectiveness and robustness of the proposed method.

Key words: Multirotors, flight control, imitation learning, reinforcement learning, robust control