Decentralized Multi-Agent Reinforcement Learning for Collision-Free Trajectory Optimization in Autonomous UAV Swarms Under GPS-Denied Environments

Operating autonomous unmanned aerial vehicle (UAV) swarms in dense, GPS-denied environments requires real-time coordination, robust dynamic obstacle avoidance, and decentralized decision-making under intermittent communication constraints. This paper proposes a distributed multi-agent deep reinforcement learning framework, SwarmOpt-MARL, designed to optimize collaborative trajectory generation and spatial formations in real time. The proposed architecture couples visual-inertial odometry state estimation with a graph attention network (GAT) to model inter-agent interactions dynamically, ensuring that each UAV can infer global swarm intent using only local sensory inputs. Hardware-in-the-loop simulations and real-world flight trials involving a 16-drone swarm across dense forest and urban canyon mockups demonstrated a 99.1% collision-free mission completion rate, alongside a 27% improvement in energy efficiency relative to classic potential-field algorithms. The results validate that distributed attention-based multi-agent policies offer high adaptability, fault tolerance, and computational scalability for search-and-rescue, reconnaissance, and disaster response operations.