Autonomous Drone Landings on an Unmanned Marine Vehicle using Deep Reinforcement Learning

Abstract

This thesis describes with the integration of an Unmanned Surface Vehicle (USV) and an Unmanned Aerial Vehicle (UAV, also commonly known as drone) in a single Multi-Agent System (MAS). In marine robotics, the advantage offered by a MAS consists of exploiting the key features of a single robot to compensate for the shortcomings in the other. In this way, a USV can serve as the landing platform to alleviate the need for a UAV to be airborne for long periods time, whilst the latter can increase the overall environmental awareness thanks to the possibility to cover large portions of the prevailing environment with a camera (or more than one) mounted on it. There are numerous potential applications in which this system can be used, such as deployment in search and rescue missions, water and coastal monitoring, and reconnaissance and force protection, to name but a few. The theory developed is of a general nature. The landing manoeuvre has been accomplished mainly identifying, through artificial vision techniques, a fiducial marker placed on a flat surface serving as a landing platform. The raison d'etre for the thesis was to propose a new solution for autonomous landing that relies solely on onboard sensors and with minimum or no communications between the vehicles. To this end, initial work solved the problem while using only data from the cameras mounted on the in-flight drone. In the situation in which the tracking of the marker is interrupted, the current position of the USV is estimated and integrated into the control commands. The limitations of classic control theory used in this approached suggested the need for a new solution that empowered the flexibility of intelligent methods, such as fuzzy logic or artificial neural networks. The recent achievements obtained by deep reinforcement learning (DRL) techniques in end-to-end control in playing the Atari video-games suite represented a fascinating while challenging new way to see and address the landing problem. Therefore, novel architectures were designed for approximating the action-value function of a Q-learning algorithm and used to map raw input observation to high-level navigation actions. In this way, the UAV learnt how to land from high latitude without any human supervision, using only low-resolution grey-scale images and with a level of accuracy and robustness. Both the approaches have been implemented on a simulated test-bed based on Gazebo simulator and the model of the Parrot AR-Drone. The solution based on DRL was further verified experimentally using the Parrot Bebop 2 in a series of trials. The outcomes demonstrate that both these innovative methods are both feasible and practicable, not only in an outdoor marine scenario but also in indoor ones as well.