Reinforcement Learning-based Fault-Tolerant Control for Unmanned Aerial Vehicles

This paper focuses on the fault-tolerant control (FTC) problem for unmanned aerial vehicles (UAVs) subject to possible multiple actuator failures, which is a tough problem to solve with traditional FTC methods since they usually require accurate mathematical models. To address the limitation of traditional FTC methods, a model-free FTC approach is proposed based on reinforcement learning (RL). Subsequently, the proposed approach is applied to construct fault-tolerant controller for UAVs without any knowledge of the quadrotor dynamic information. Then, an end-to-end control policy that can tolerant actuator failures is obtained, which can map the state of the UAVs directly to the control commands of the four rotors after learning. Finally, the effectiveness of the proposed fault-tolerant approach is demonstrated by using the flexible modular quadrotor simulator.