Research on Attitude Control of Launch Vehicle at Ascent Stage Through Reinforcement Learning

Focusing on the problem that existing control methods of launch vehicle have insufficient adaptability to the uncertainty in actual flight, an attitude control method of launch vehicle at ascent stage through reinforcement learning (RL) is proposed in this paper. Firstly, a public recorded launch vehicle named Ares-I is taken as the research object, and its dynamic model is built in the specific reference frame. Then, an RL framework based on Soft Actor Critic (SAC) algorithm is constructed for attitude control of Ares-I and a reward function considering the tracking error, trajectory stability and training speed is designed. Finally, through offline training, an RL based attitude controller of Ares-I is proposed. The simulations show that this method improves the control accuracy by an order of magnitude compared with the traditional PID controller.