Deep-Reinforcement-Learning-Based Collision Avoidance in UAV Environment

Abstract

Unmanned Aerial Vehicles (UAVs) have recently \nattracted both academia and industry representatives due to \ntheir utilization in tremendous emerging applications. Most \nUAV applications adopt Visual Line of Sight (VLOS) due to \nongoing regulations. There is a consensus between industry for \nextending UAVs’ commercial operations to cover the urban and \npopulated area controlled airspace Beyond VLOS (BVLOS). \nThere is ongoing regulation for enabling BVLOS UAV management. Regrettably, this comes with unavoidable challenges \nrelated to UAVs’ autonomy for detecting and avoiding static \nand mobile objects. An intelligent component should either \nbe deployed onboard the UAV or at a Multi-Access Edge \nComputing (MEC) that can read the gathered data from \ndifferent UAV’s sensors, process them, and then make the \nright decision to detect and avoid the physical collision. The \nsensing data should be collected using various sensors but \nnot limited to Lidar, depth camera, video, or ultrasonic. This \npaper proposes probabilistic and Deep Reinforcement Learning \n(DRL)-based algorithms for avoiding collisions while saving \nenergy consumption. The proposed algorithms can be either run \non top of the UAV or at the MEC according to the UAV capacity \nand the task overhead. We have designed and developed \nour algorithms to work for any environment without a need \nfor any prior knowledge. The proposed solutions have been \nevaluated in a harsh environment that consists of many UAVs \nmoving randomly in a small area without any correlation. The \nobtained results demonstrated the efficiency of these solutions \nfor avoiding the collision while saving energy consumption in \nfamiliar and unfamiliar environments.