Physics-prior-driven Distributed Deep Reinforcement Learning for Multi-Uav Path Planning

Coordinating multi-UAV path planning in dynamic environments suffers from severe online computational bottlenecks and non-stationarity. To address these challenges, we propose a physics-prior-driven decentralized deep reinforcement learning (DRL) framework. Functioning as a scalable distributed computing paradigm via decentralized training with decentralized execution (DTDE), the framework mitigates the curse of dimensionality. First, an improved artificial potential field (APF) translates three-dimensional (3D) spatial forces into deterministic control priors, continuously blended with a multi-agent soft actor–critic (MASAC) policy to accelerate offline convergence and save high-performance computing (HPC) resources. Second, embedded long short-term memory (LSTM) networks process locally augmented states for implicit neighbor intent inference without communication overhead. Finally, by integrating these components into the unified APF-LSTM-MASAC architecture, we shift computational burdens to offline HPC training, enabling real-time online inference at the edge. Compared to baseline DRL algorithms, it significantly improves trajectory smoothness and reduces energy consumption by 15.04