Reinforcement Learning Applied to a Quadrotor Guidance Law in Autonomous Flight

摘要

Autonomous flight of Unmanned Aerial Vehicles (UAVs) in unknown or uncertain environments can benefit from control methods that are able to learn and adapt to these conditions. This paper presents the setup and results of a high level reinforcement learning problem for both simulation and real flight tests. The problem provided is that of a quadrotor taking pictures of a disaster site. The environment is completely unknown at first and the agent must learn where the sites of interest are and the most efficient way to get there. The results show that the quadrotor agent can learn a converged, near optimal value function after many iterations. However, a non-converged value function can result in the same desirable actions with much fewer iterations. Furthermore in this paper, a research and test laboratory for ground robots and aerial vehicles is presented.