Adaptive Multi-Agent UAV Path Planning: From Classical Optimization to Deep Reinforcement Learning

Abstract:Multi-agent UAV path planning in three-dimensional environ- ments with obstacles is a problem of fundamental importance for search-and-rescue, surveillance, and logistics operations. Existing classical optimization methods produce high-quality paths for static environments but fail when obstacles move or missions change mid-flight. We present a systematic progres- sion from a Grey Wolf Optimizer (GWO) baseline to a Deep Reinforcement Learning (DRL) framework, training both DQN and PPO agents in a custom OpenAI Gym environment for 3D UAV navigation. Our key finding: PPO converges 40% faster than DQN under sparse-reward conditions—the realistic setting where agents receive feedback only upon goal completion or collision. We benchmark all three methods on identical test environments, measuring convergence speed, path optimality, and computational cost. The results reveal a fundamental trade- off: classical methods require zero training but cannot adapt; DRL requires upfront training but enables real-time obstacle avoidance. PPO offers the best cost-quality ratio, making it the preferred choice for practical UAV swarm deployments.