Unmanned active observation planning of near-surface ocean under tropical cyclone condition

Studying tropical cyclone (TC)-ocean interactions is of profound significance for advancing our understanding of these extreme weather systems. However, conventional observational methods face substantial challenges in capturing these dynamic interactions for further high-precision quantitative analyzes. Recent studies demonstrate that autonomous underwater vehicles (AUVs) possess the unique capability to operate within TC-induced shallow-layer flow fields. Capitalizing on this distinctive advantage, we establish a novel framework for unmanned active observation planning of tropical cyclones using AUV swarms. Building upon our previous work on estimating TC centers from subsurface flow fields, the cornerstone of this framework lies in a multi-agent task planning strategy that enables each AUV to actively prioritize observation targets while balancing energy consumption, task execution time, and data redundancy minimization. Specifically, the planning algorithm integrates game theory with reinforcement learning (RL), employing continuous action space methods (SAC, PPO, TD3) in the inner loop, with proven convergence to Nash equilibrium. Drawing upon real-world data from TC Hinnamnor (2022) and six additional representative TC cases, we demonstrate that while traditional methods struggle to balance individual utility with global efficiency in TC observation, our game-theoretic approach effectively resolves the utility balancing problem in multi-agent decision-making while demonstrating strong robustness across diverse TC scenarios and inner-loop RL algorithms. More significantly, this work offers not only a novel solution for the task planning in TC observation but also establishes a versatile framework with broad applicability to complex multi-agent scenarios.