Attitude Controller Design based on Deep Reinforcement Learning for Low-cost Aircraft

In this paper, for a class of low-cost aircraft whose accurate mathematical models can't be obtained, a kind of attitude controller design method based on deep reinforcement learning is proposed. Considering the state and control signals in attitude control are continuous variable, an anthropomorphic attitude controller design method based on the deep deterministic policy gradient (DDPG) theory is researched. Based on the actor-critic structure, the controller can be learned through continuous interaction and trial between the controller and the system environment, which reduces the dependence on the accurate aircraft model. The simulation results show that the DDPG based attitude controller can achieve a good control accuracy, and its end-to-end design process can effectively reduce the design complexity.