Vision-Based Autonomous Landing of a Multi-Copter Unmanned Aerial Vehicle using Reinforcement Learning

This paper presents vision-based landing guidance of multi-copter Unmanned Aerial Vehicle (UAV) using reinforcement learning. In this approach, the guidance method is not designed or proposed by a human, but deployed by a neural network trained in simulated environments; which contains a quad-copter UAV model with Proportional-Integral-Derivative (PID) Controller, ground looking camera model that gives pixel deviation of targeting landing location from the center of an image frame, and laser rangefinder that gives altitude above ground level. Since we aimed for various types of multi-copter UAVs to track targeting ground location, reinforcement learning method has been used to generate proper roll and pitch attitude commands in multiple situations. Series of flight experiments show that a multi-copter UAV equipped with a proper attitude controller and trained artificial intelligence pilot can guide a multi-copter UAV to a ground target position.