Unmanned Aerial Vehicle Trajectory Planning via Staged Reinforcement Learning

Unmanned Aerial Vehicle (UAV) trajectory planning problem has always been a popular but still an open topic, where online planning is desired in unknown environments. This paper investigates how to combine human knowledge with reinforcement learning to train the UAV in a staged manner. With the novel framework we design, the UAV learns well to avoid densely arranged no-fly-zones and reach stationary or moving targets via calling the trained policy online. We demonstrate the advantages of our approach in terms of the flight time and the success rate of reaching target and avoiding no-fly-zones. The experimental results are performed in a set of new designed environments including dynamic no-fly-zones and moving targets.