Reliable Collision Avoidance for Small UAVs via Deep Reinforcement Learning with Unified Data Representation

Collision avoidance is a fundamental capability for multi-UAV systems to complete autonomous tasks. For small rotary-wing UAVs, due to their limited perception capacities and flight endurance, achieving reliable collision avoidance remains challenging. To this end, we propose a reliable multi-UAV collision avoidance method based on deep reinforcement learning, which leverages information from neighboring UAVs to augment the onboard perception capabilities. This method utilizes a technique called unified data representation (UDR) to combine depth data from LiDAR measurements and information from neighboring UAVs into a compact feature embedding, which can effectively integrate various observations with significant dimensional differences and inconsistent environmental expressions. Based on this, the collision avoidance policy hosted by a neural network is learned and produces control references for each UAV to ensure safety during flight. The training experiments demonstrate that our method achieves approximately a 20% improvement in average rewards compared to baseline approaches, using the same number of training episodes. Additionally, we perform validation experiments using the PX4 flight controller in a soft-in-the-loop (SITL) setup within the Gazebo simulator. Experimental results show that our approach significantly improves the reliability of collision avoidance, particularly in scenarios involving inter-aircraft collisions.