Reinforcement Learning Based Assistive Collision Avoidance for Fixed-Wing Unmanned Aerial Vehicles

In recent years, Unmanned Aerial Vehicles (UAV) have assumed the role of a key technological component for a variety of different applications. Many missions will require integration into civil airspace and Beyond Visual Line of Sight (BVLOS) operations. To make BVLOS mission possible is necessary to guarantee a level of safety equal to the case of Visual Line of Sight operations. Non-cooperative entities in the airspace must be considered. Provided that such entities are detected and tracked, efficient and robust avoidance manoeuvres are required. Reinforcement Learning methods may improve the efficiency and robustness of avoidance manoeuvres, as they are potentially more adaptable to complex and unseen situations than classical approaches. This work aims to present a proof-of-concept for a reinforcement-trained collision avoidance system for a fixed-wing UAV. A novel geometric-based logic for conflict representation has been defined. Based on this novel logic, different agents have been trained to avoid random mid-air collisions with an uncooperative intruder. The trained agents have been subsequently validated and compared with thousands of randomly generated conflicts.