Fault-tolerant control for UAV based on deep reinforcement learning under single rotor failure

In order to achieve high-speed flight of a quadrotor UAV when one single rotor completely fails, a fault-tolerant control method based on deep reinforcement learning was designed. Since our policy is a low-level control, neural networks are used in reinforcement learning to directly map states to motor action values. The learning algorithm is developed based on the proximal policy optimization algorithm. By introducing PID lift correction into the actor-critic structure, both control accuracy and robustness are significantly improved. The proposed method has been tested in a flight simulator, and the results show that the reinforcement learning strategy is highly robust to model uncertainty and external interference.