Deep Reinforcement Learning-Based Physical Verification of Autonomous Landing for UAV

In recent years, applying deep reinforcement learning (DRL) method to UAV (Unmanned Aerial Vehicle) field has gained more and more attention, including but not limited to swarm control, mission planning, obstacle avoidance, etc. Due to the trial-and-error mechanism of the DRL method, many existing methods, which work well in simulations, may be hard to be used directly in the real system. In this paper, the physical verification experiments of autonomous landing for UAV are achieved through the DRL method. An improved DRL method based on deep Q-learning, namely dueling-double-deep Q-learning (3DQN) algorithm, is introduced. Besides, an accelerated training method for the proposed DRL algorithm is presented, which can not only speed up the convergence of the DRL network, but also eliminate the gap between simulation training and physical experiment in a certain degree. The real flight experiment validates the effectiveness and reasonableness of the proposed landing algorithm based on DRL.