Experimental Verification of Deep Reinforcement Learning Attitude Controller

Unmanned Aerial Vehicle (UAV) attitude control faces challenges in nonlinear dynamics, environmental disturbances, and actuator constraints. This study proposes a Deep Deterministic Policy Gradient (DDPG)-based intelligent control framework for end-to-end UAV attitude regulation. A compound reward function integrating attitude error, angular rate penalty, and actuator constraints is designed to enhance robustness. The simulation results show that the airspeed tracking accuracy of the controller is less than 0.5 m/s during the horizontal flight, climbing, sliding and coordinated turning tasks, and the dynamic response time of the attitude angle is less than 0.5 s. Real-world experiments on a Skywalker-X8 fixed-wing UAV, equipped with an NVIDIA processor, demonstrate real-time control with average roll and pitch errors of 3.0° and 2.3°, respectively. The framework’s feasibility is confirmed under varying flight conditions, providing a novel paradigm for intelligent flight control engineering.