Abstract
This paper presents a novel reinforcement learning (RL) approach for a tethered drone to follow a predefined three-dimensional trajectory in a GPS-denied environment. The adopted Q-learning strategy determines high-level actions using raw observations from the onoard accelerometers, gyros, and altimeter, which facilitates a low-level proportional-integral-derivative (PID) controller to drive the drone through the desired waypoints on a reference trajectory. The effectiveness of the proposed approach is demonstrated in a simulated environment.