Deep Reinforcement Learning Control for Aerobatic Maneuvering of Agile Fixed-Wing Aircraft

Autonomous aerobatic flight has been a challenging control problem for many years. This is because aerobatic flight requires such highly precise control while operating on the extreme edges of the flight envelope which most controllers are not able to handle. For pilots, this precise control has been learnt through many years of flight experience. The research in this paper significantly shortens this learning time by extending the state of the art work in Deep Reinforcement Learning to the realm of flight control. This paper presents a Normalized Advantage Function controller that, unlike traditional controller architectures, emulates stickto- surface control by utilizing a simple unified nonlinear control framework to command and coordinate full multi-actuator authority. Through trial-and-error simulated experiences, the controller is able to explore the full range of the nonlinear flight envelope and learn by itself, without human input, an aerobatic maneuver in little time on the order of hours. This controller utilizes the large multidimensional state and action spaces of the aircraft to optimize aerobatic performance and high degree autonomous flight skill. We are able to showthrough high-fidelity simulations that this controller is able to successfully learn and execute two different aerobatic maneuvers: Slow Roll and Knife Edge.