Imitation Learning for Neural Network Autopilot in Fixed-Wing Unmanned Aerial Systems

This research identifies the feasibility of training an artificial neural network (ANN) autopilot, using supervised learning techniques, including an imitation learning framework known as the data aggregation set (DAgger) algorithm. The ANN autopilot aims to mimic a unified guidance, navigation & control (GNC) system to fly a fixed-wing UAS. Utilizing high-fidelity nonlinear 6-DOF aircraft simulations, it is shown that several modifications in existing imitation learning techniques must be considered. DAgger algorithm when applied sequentially to augment data along the desired flight trajectory to train the ANN autopilot is unable to generalize across turning and straight-line maneuvers, and hence cannot learn to fly stably, eventually terminating the simulation. Monte-Carlo methods when incorporated into the DAgger algorithm allow for sampling of random data along the desired flight trajectory, proved to be effective to train the ANN autopilot to fly indefinitely with acceptable tracking errors, however undesirable low frequency oscillations were observed in control input and aircraft states. The oscillatory behavior is dealt with by introducing a time-based moving window along the trajectory for data addition in conjunction with the standard DAgger algorithm. Different variations of DAgger algorithms are tested and compared using closed-loop 6-DOF flight simulations. 3D trajectory tracking and stability in aircraft states are evaluated with evidence supporting the idea that an ANN autopilot can behave as a unified GNC system and fly a fixed-wing UAS.