Deep Reinforcement Learning Attitude Control of Fixed-Wing UAVs

Abstract

The fixed-wing UAV is a non-linear and strongly coupled system. Controlling UAV attitude stability is the basis for ensuring flight safety and performing tasks successfully. The non-linear characteristic of the UAV is the main reason for the difficulty of attitude stabilization. Deep reinforcement learning for the UAV attitude control is a new method to design controller. The algorithm learns the nonlinear characteristics of the system from the training data. Due to the good performance, the PPO algorithm is the mainly algorithm of reinforcement learning. The PPO algorithm interacts with the reinforcement learning training environment by gazebo, and improve attitude controller, different from the traditional PID control method, the attitude controller based on deep reinforcement learning uses the neural network to generate control signals and controls the rotation of rudder directly.