Developing Flight Control Policy Using Deep Deterministic Policy Gradient

Developing a flight control system for a 6 degree-of-freedom aircraft remains a considerable task that requires time and effort to gather all the necessary data. In this paper, a policy using reinforcement learning based on Deep Deterministic Policy Gradient (DDPG) is proposed and its application to UAS (Unmanned Aerial System) control is presented. Previous research has shown a slight difficulty in training the DDPG learning agent for a system with multiple agent. A learning strategy is introduced to implicitly guide the learning agent to utilize all three control surfaces and still produce a converging policy. The DDPG learning agent is trained through several training sets to generate the best policy suited to control the aircraft. The final policy as the result of the training procedure is then extracted and tested. This research shows that DDPG can be used to develop the policy for flight control.