Fighter Pilot Behavior Cloning
Viktor Sandstrom, Linus J. Luotsinen, Daniel Oskarsson
Abstract
In this paper, a feed-forward neural network is trained on a small dataset of human fighter pilot data, recorded from maneuvering a fixed-wing fighter aircraft in a flight simulator. The goal is to model the pilot behavior, using a technique called behavior cloning. By carefully preprocessing the training data, it is shown that this simple and intuitive approach results in a model that can successfully fly the aircraft at high velocity on flight tracks that demand sharp turns, and even perform maneuvers not explicitly represented in the data. Furthermore, it is demonstrated that a pretrained neural network will adapt to a significant change in flight dynamics with less training, compared to a previously untrained model. This transfer learning scenario is important since fine-tuning pretrained models could simplify the development of a wide fleet of AI aircraft.