Fighter pilot behavior cloning and transferring to another aircraft

Abstract

he "Pilot-in-the-loop" flight simulators are important tools in the development of fighter aircraft because they allow engineers and designers to test different scenarios and algorithms in a controlled environment without the risks associated with actual flight testing. However, these simulations often require high pilot skill levels and can be time-consuming and costly to arrange. It is important to create realistic models of human fighter pilots in order to reduce the reliance on skilled pilots to demonstrate agile/aerobatic maneuvers in flight simulators. Traditional controllers for aircraft require detailed knowledge about the aerodynamic model and physics of the aircraft in order to perform aerobatic maneuvers. Also, these control algorithms may not be able to match the performance of skilled human pilots, who are limited by a lack of bandwidth. This suggests that there may be potential for improving the performance of aircraft through the use of techniques that can take advantage of the superior speed and maneuverability of skilled pilots. In that case, imitation learning is a potential solution to eliminate the dependency need of skillful pilots in the flight simulator. Imitation learning also known as learning from demonstrations has benefited from computational progresses brought on by deep learning and increased availability of demonstration data. It is aimed to emulate desired behavior in a given task. An agent is trained to learn mapping between observations and actions by utilizing demonstrations. In this thesis aims the development of a pilot behavior model which is capable of autonomously performing agile maneuvers and is able to replace expert pilots' demonstrations over its full flight envelope in the flight simulator. Moreover, this model is transferable to other aircraft with limited data using transfer learning techniques. Besides all these features, the pilot behavior model can be able to run in real time in the flight simulator.