GADCA: A MARL-Based Graph-Attentive Decentralized Collision Avoidance Framework for UAV in UTM

With increasing congestion in urban airspace, centralized Unmanned Aircraft System Traffic Management (UTM) face scalability challenges and risks of single-point failures, driving the need for robust decentralized solutions to ensure safe autonomous UAV navigation. This paper introduces the Graph-Attentive Deep Collision Avoidance (GADCA) framework, a decentralized Multi-Agent Reinforcement Learning (MARL) approach for collision avoidance in urban airspace. GADCA extends the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) by integrating Graph Attention Networks (GAT) to model dynamic inter-agent relationships, enabling UAVs to navigate effectively using local observations and shared situational awareness. Addressing limitations in prior work, GADCA incorporates 3D dynamics, continuous velocity action spaces, and scalable fleet sizes for robust performance. Unlike centralized UTM methods, GADCA operates in a decentralized manner during execution, where each UAV independently makes collision-avoidance decisions based on local and shared information, without relying on a central controller. It combines decentralized MARL execution with graph-based relational reasoning to enhance collision avoidance. Experimental results show that GADCA outperforms MADDPG across fleets of 5 to 10 UAVs, achieving 45.8% to 67.7% improvements in learning performance, 12.0% to 54.5% enhancements in task completion, up to 23% reduction in conflicts, and 58% to 100% reduction in collision risk, with no physical collisions recorded. The framework’s dynamic adjacency matrix and collective reward mechanism effectively balance navigation efficiency and safety in complex 3D urban environments, using a highly realistic Gauss-Markov mobility model.