Deep Reinforcement Learning Automatic Landing Control of Fixed-Wing Aircraft Using Deep Deterministic Policy Gradient

Landing phase remains to be one of the most crucial and difficult tasks to achieve among the flight envelope of an aircraft. The proof-of-concept controller in this research implemented the use of DDPG (Deep Deterministic Policy Gradient), a DRL (Deep Reinforcement Learning) approach in attempt to find policies of aircraft landings given designed requirements, or rewards. This research provided new methods in reward shaping, or reward engineering used during training and the investigation of the effects of hyperparameters and different network topologies of Neural Networks in training of aircraft landing control. The results of outer loop control and direct control in this research using DDPG on aircraft landing, with comparisons of numerous baseline and Neural Network approaches, proves the ability and potential of such DRL method, and is validated in numerous wind disturbance conditions, which demonstrated the robustness of DDPG agent. It is also found that besides the ability of DDPG agents to develop control policies for aircraft landings, such method provides insights of the controls and states of aircraft during landing, enabling guidelines of the flight characteristics of the aircraft in landing for pilots or design of controllers.