A Model-free Flat Spin Recovery Scheme for Miniature Fixed-wing Unmanned Aerial Vehicle

The present paper proposes a Deep- Reinforcement-Learning-based (DRL-based) model-free flat spin recovery scheme to recover a miniature unmanned aerial vehicle (UAV) back to steady level flight swiftly. Two types of deep reinforcement learning (DRL) are utilized for the two recovery phases to fully exploit DRL’s strengths in model-free situations. In the first phase, the angular rates of UAV are attenuated swiftly by a deep Q-network (DQN); in the second phase, the UAV is continuously regulated with a novel algorithm termed LA-DDPG, which is the deep deterministic policy gradient (DDPG) exploring by learning automata (LA). Simulation tests show that the proposed recovery scheme can recover the UAV from flat spin mode to steady level flight with satisfactory performance without relying on accurate prior models or other control structures. Furthermore, LA-DDPG exhibits better stability and efficiency than DDPG.