Dynamic UAV Path Planning Based on Reinforcement Learning With Posterior Zone Priority Weighted-GMM

Efficient UAV path planning in large and dynamic environments requires adaptive strategies that account for varying regional importance. Traditional coverage or Reinforcement Learning-based methods often suffer from redundant exploration and unstable learning, especially as the swarm size increases. This paper proposes a reinforcement learning framework enhanced by a Posterior Zone Priority Weighted Gaussian Mixture Model (PZP–GMM), which fuses posterior information gain, intrinsic priority, and spatial proximity into a time-varying probabilistic field. This model provides structured guidance that improves exploration and reduces motion inefficiency. For single-UAV planning, PZP–GMM is embedded into a discrete 3D RL formulation to drive entropy-based, priority-aware exploration. For multi-UAV coordination, we introduce a Critical-Zone-Aware (CZA) assignment mechanism and a Joint Priority–weighted PPO (JP–PPO) algorithm that applies priority-gated advantages and class-conditioned entropy within a Centralized-Training–Decentralized-Execution (CTDE) framework. Experimental results show that PZP–GMM improves reward stability, coverage, and trajectory coherence for a single UAV. In multi-UAV deployments, the proposed JP–PPO with PZP–GMM guidance achieves consistently high coverage, reaching up to 97% coverage, while reducing redundant visitation and improving mission efficiency. These results demonstrate a scalable and priority-aware solution for dynamic UAV path planning.