Reinforcement Learning Training Environment for Fixed Wing UAV Collision Avoidance

In recent years, Unmanned Aerial Vehicles (UAVs) are emerging as a key technology with possible uses for several applications. To fully exploit their potential, UAVs needs to be able to be fully automatized and they need to be allowed to go Beyond Visual Line of Sight (BVLOS). Nowadays, autopilots and Detect and Avoid (DAA) systems, essential components for BVLOS operations, are developed with classical control approaches and classical algorithm for identification and avoidance of collision in the airspace. However, new methods based on Artificial Intelligence and, in particular, on Reinforcement Learning are attracting scientific interest. In order to train an agent with Reinforcement Learning methods, the development of the training environment is of utmost importance. This paper aims to present the development of a training environment to train an agent to avoid collisions in the airspace by setting desired values of roll, pitch and throttle. It will describe the developed training environment and the PID controllers implemented for allowing the own aircraft to receive the roll, pitch and throttle commands and to control the intruders' trajectory.