Three‐Dimensional Trajectory Tracking Control for Stratospheric Airship Based on Deep Reinforcement Learning

The stratospheric airship is a low‐speed aircraft capable of remaining in near space for extended periods. It can transport payloads to perform various tasks, including intelligence gathering, reconnaissance, and surveillance. Trajectory tracking control is a critical technology that enables the airship to execute its missions, such as relocating flight sites and maintaining station‐keeping. Focusing on the large error in three‐dimensional trajectory decoupling control of stratospheric airship, a three‐dimensional trajectory direct tracking control method based on reinforcement learning was researched. According to the discrete state and action space, a Markov decision process model for the three‐dimensional trajectory tracking control problem was established. The Boltzmann random distribution of reward value and probability of wind direction angle were taken as the action selection criteria of the Q ‐learning algorithm, the cerebellar model articulation controller (CMAC) neural network was constructed for the discrete action value, and the optimal action sequence was fast obtained. Taking the cube and spherical discrete state distribution as an example, the trajectory tracking control method was validated through simulation. Simulation results show that, by properly setting the action range of actuators including fans, valves, and thrust vectors and refining state action value distribution, the proposed three‐dimensional trajectory control method has high tracking accuracy, the horizontal error is in a 10‐m level, the altitude error can achieve meter level, and the tracking trajectory is smooth and has well engineering realizability.