Monocular Vision based Autonomous Landing of Quadrotor through Deep Reinforcement Learning

An improved deep reinforcement learning (DRL) method is proposed to solve autonomous landing problem of quadrotor. Autonomous landing is a significant function for unmanned aerial vehicle (UAV) such as quadrotor. Previous solutions are mainly based on relative position calculation or the landmark detection, which either needs massive additional sensors or lacks intelligence. In this paper, we focus on realizing autonomous landing through DRL method. Whole landing process is implemented by an improved deep Q-learning network (DQN) based end-to-end control scheme. Only one down-looking camera is used to capture raw images directly as input states. An Aruco tag is placed at the landing region for feature extraction. Double network and the dueling architecture are applied to improve DQN algorithm. Besides, the reward function is well designed to fit the auto-landing scenario. The experiments show that the improved DQN can make the quadrotor land on the landmark successfully and achieve better performance while comparing to the original deep Q-learning solution.