Dynamic Weather Avoidance Based on Maximum Diffusion Reinforcement Learning

This paper investigates an autonomous trajectory replanning framework for aircraft operations based on Maximum Diffusion Reinforcement Learning. The method fully considers the complex coupling between aircraft high-inertia kinematics and hazardous weather obstacles, addressing the critical exploration failure challenge inherent in conventional Deep Reinforcement Learning (DRL) algorithms. Specifically, high-fidelity flight dynamics based on the Base of Aircraft Data (BADA) model are integrated into the decision-making process. To overcome the exploration limitations of standard Maximum Entropy RL caused by strong temporal correlations in state transitions, a trajectory entropy maximization mechanism is introduced. This mechanism utilizes the determinant of the trajectory autocovariance matrix as a regularization term to enforce diverse state-space coverage. Numerical results and ablation studies demonstrate that the proposed framework significantly outperforms the baseline Soft Actor-Critic (SAC) algorithm. In scenarios with complex obstacle distributions, the proposed method achieves and higher collision avoidance success rates. Furthermore, the ablation analysis verifies the critical role of the diffusion regularization term in preventing policy collapse and ensuring robust dynamic replanning capabilities.