Autonomous Flight Arcade: Reinforcement Learning for End-to-End Control of Fixed-Wing Aircraft

Abstract

In this paper, we present the Autonomous Flight Arcade (AFA), a suite of robust environments for end-to-end control of fixed-wing aircraft and quadcopter drones. These environments are playable by both humans and artificial agents, making them useful for varied tasks including reinforcement learning, imitation learning, and human experiments. Additionally, we show that interpretable policies can be learned through the Neural Circuit Policy architecture on these environments. Finally, we present baselines of both human and AI performance on the Autonomous Flight Arcade environments.