A Deep Reinforcement Learning-Based Approach for Autonomous Quadrotor Tracking Control

Autonomously tracking a dynamic unmanned ground vehicle (UGV) with an unmanned aerial vehicle (UAV) is challenging due to the difficulty for the UAV to track the UGV’s real-time state and adjust its policy accordingly. In this paper, we propose a deep reinforcement learning approach for a quadrotor UAV to track a moving UGV without knowing its motion dynamics. The proposed learning framework consists of two main systems: a Tracking Vision System (TVS) utilizing deep learning and a Tracking Control System (TCS) utilizing deep reinforcement learning. The TVS recognizes the UGV by analyzing images captured by the UAV, and then it provides the TCS with input data that is used to fulfill Markov Decision Process (MDP) in the TCS. The TCS generates optimal landing policies learned through simulated training. Moreover, we propose an Expert Knowledge Transfer (EKT) algorithm that allows human intervention in the UAV’s action space, which can speed up the training process. We demonstrate the effectiveness of the proposed method in simulation experiments.