Adaptive curriculum learning-based path planning and data collection for autonomous underwater vehicles under extreme sea conditions

Abstract

Path planning and data collection for Autonomous Underwater Vehicles (AUVs) under extreme sea conditions face severe environmental disturbances and high task complexity, leading to low training efficiency and divergence in traditional reinforcement learning. To address the limitations of existing fixed curriculum learning (CL) methods—namely manual stage presetting, tedious hyperparameter tuning, and restricted generalization—this paper proposes an adaptive CL framework tailored for partial differential equation (PDE)-coupled ocean environments. First, a multi-dimensional difficulty parameterization system driven by PDEs is designed to achieve graded environmental complexity variations by dynamically scaling wave amplitude and turbulence intensity. Second, an adaptive transition criterion based on a sliding-window success rate is introduced to dynamically advance difficulty according to the agent's real-time learning progress. Simulation experiments based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm demonstrate that in complex tasks, the proposed method significantly enhances data collection volume by 35.8% (Holm-corrected p < 0.01) and reduces collisions by 44.9%. Generalization tests confirm that the method retains the best performance on all metrics at up to three times the training difficulty, demonstrating superior robustness. This study provides an efficient training paradigm for agents in complex fluid dynamics environments.