A Deep Reinforcement Learning Approach for UAV Collision Avoidance in Integrated Airspace

The increasing presence of Unmanned Aerial Vehicles (UAVs) in integrated airspace raises collision risks with manned aircraft. To address this challenge, we propose a two-layer collision avoidance model consisting of a warning zone and an avoidance zone, specifically designed for UAVs operating in complex airspace. Using an advanced variant of the Deep Q-Network (DQN) called D3QN, we trained UAVs to adaptively respond to varying conditions, including dynamic wind patterns. Simulations conducted in a 30 km × 30 km integrated airspace demonstrate that even under highest wind level, D3QN achieves a 94.7% success rate in collision avoidance, significantly surpassing the 88% success rate of the DQN model. Beyond achieving higher success rates, D3QN also converges faster, demonstrating robustness under fluctuating conditions. This study provides key insights into enhancing UAV collision avoidance, supporting the safe and efficient integration of UAVs into shared airspace.