AirGuardian : a parallel autonomy approach to self-flying planes

Abstract

In this paper we lay the foundation for a fixed-wing parallel autonomy aircraft system in which both the autonomous component of the aircraft and the pilot jointly fly the plane resulting in an overall safer and more effective flying experience. Experimenting with both imitation learning and reinforcement learning we develop the fully autonomous flight component of this system through the case study of flight within a canyon. Our strategy includes vision-based learning with both camera and depth map inputs to create an end-to-end learning platform for aircraft control. Using a yaw based controller and reinforcement learning, we are able to demonstrate stable flight in a unknown canyon environments that exhibit unexpected hazards such as wind gusts and terrain changes.