A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform

Abstract

The use of multi-rotor UAVs in industrial and civil applications has been extensively encouraged by the rapid innovation in all the technologies involved. In particular, deep learning techniques for motion control have recently taken a major qualitative step, since the successful application of Deep Q-Learning to the continuous action domain in Atari-like games. Based on these ideas, Deep Deterministic Policy Gradients (DDPG) algorithm was able to provide outstanding results with continuous state and action domains, which are a requirement in most of the robotics-related tasks. In this context, the research community is lacking the integration of realistic simulation systems with the reinforcement learning paradigm, enabling the application of deep reinforcement learning algorithms to the robotics field. In this paper, a versatile Gazebo-based reinforcement learning framework has been designed and validated with a continuous UAV landing task. The UAV landing maneuver on a moving platform has been solved by means of the novel DDPG algorithm, which has been integrated in our reinforcement learning framework. Several experiments have been performed in a wide variety of conditions for both simulated and real flights, demonstrating the generality of the approach. As an indirect result, a powerful work flow for robotics has been validated, where robots can learn in simulation and perform properly in real operation environments. To the best of the authors knowledge, this is the first work that addresses the continuous UAV landing maneuver on a moving platform by means of a state-of-the-art deep reinforcement learning algorithm, trained in simulation and tested in real flights.