Research On UAV Autonomous Landing Method Based On Deep Reinforcement Learning

In order to ensure the ease of operation of the drone and the goal of liberating human,and to improve the level of autonomy required to provide multi-rotor drones with long-term missions,the drone’s autonomous landing and homing operations on the recovery platform after the completion of the mission has become a hot issuse of research.Since most of the recovery platforms are maneuverable,autonomous landing on the mobile platform has become an urgent problem to be solved.The main purpose of this paper is to control the drone through the machine learning method to solve its landing problem on the ground mobile platform.Aiming at the problem of poor repeatability and high risk of drone learning and training in the real world,using Parrot’s Bebop2 as a drone unit and using the Parrot company’s Sphinx simulator,a set of Gazebo-Parrot and ROS-based UAV simulation system is established.This system runs under Ubuntu/Linux system.Through this system,the learning and training of the UAV in the simulation environment is realized,and the problems of poor repeatability and high risk of physical training are solved.This paper also elaborated on the development background of deep reinforcement learning algorithms and the specific principles together with specific details of some of the algorithms,analyzed the advantages and disadvantages of each algorithm and how to select and apply deep reinforcement learning algorithms on the UAV’s specific mission of landing on the mobile platform.Aiming at the mutual perception of platforms and drones,this paper proposes two feasible solutions: external assistance and machine vision.Among them,the external assistant is replaced by the Gazebo simulator in the simulation,and the machine vision scheme is theoretically demonstrated and explained in detail to solve the problem that there may be no external equipment in the real world to provide landing-related data.At the same time,according to the different mobile conditions of the platform,this paper designs the mobile platform with 4 different state conditions and designs a traditional controller to compare with the machine learning algorithm.Due to the high level of simulation and the complete physics engine of the Gazebo platform,this solution also has high portability for real-world operations.