Autonomous Landing on a Moving Platform Using Vision-Based Deep Reinforcement Learning

This letter describes autonomous landing of an unmanned aircraft system on a moving platform using vision and deep reinforcement learning. Landing on the moving platform offers several benefits, such as more mission flexibility and reduced flight time. In particular, the end-to-end vision approach (i.e., an input to the reinforcement learning is a raw image from the camera) with the deep regularized Q algorithm and custom designed reward is utilized. The custom reward was specifically devised to encourage useful feature extraction from the state space. Additionally, the proposed reinforcement learning algorithm has full 3D velocity control including the vertical channel. The simulation results show that the proposed approach can outperform existing approaches which use high-level extracted features (such as relative position and velocity of the landing pad). The simulation results are then successfully transferred to the real-world experiment by utilizing domain randomization.