Sim-to-Real Deep Q-Learning with Human-Aided Demonstrations for Vision-Based Precision Landing of Multirotors

Enabling the capability of accurate and reliable pre-cision landing for multirotor using deep reinforcement learning is presented in this paper. A realistic quadrotor landing environ-ment was developed using Gazebo, taking wind disturbance into consideration, in order to train a robust control using the Deep Q-Network algorithm. Human-aided demonstrations are presented in order to increase learning speed. It navigates the quadrotor to the landing site through vision-based methods, which include AprilTags detection. We show the feasibility of a smooth sim-to-real transfer with minimum adjustments through extensive simulations and real-world experiments, showing that the landing error is less than 5 cm in real-world tests.