Acquisition of Hovering by Actual UAV Using Reinforcement Learning

In this study, we applied reinforcement learning to actual quadrotor unmanned aerial vehicles (UAVs). It is expected that UAVs will contribute to multiple fields (e.g., rescue, sports, and entertainment). However, the autonomous control of UAVs is still a difficult problem to solve. We applied reinforcement learning to a UAV to achieve stable hovering. Q-learning, a common reinforcement learning method, was used in our study. Several previous studies on learning controllers for UAVs have been conducted. However, most of these studies only carried out computer simulations to verify the effectiveness of the learning method. Conversely, we conducted our experiment using an actual quadrotor UAV. The experimental results demonstrate that the UAV can acquire knowledge to achieve stable hovering over a marker placed on the ground. We also compared the performances of a trained UAV (using the learning method) and a UAV controlled by a PID controller. The result indicates that a trained UAV can hover as stably as one controlled by the PID.