Cooperative Control in Bottleneck Scenarios for Dynamically Networked USVs Using Graph Attention and Weighted Mean-Field Approximation

Abstract

With the rapid evolution of autonomous shipping hubs, high-frequency logistics operations increasingly require multiple independent uncrewed surface vehicles (USVs) to navigate through confined waterways. In such scenarios, multi-vessel interactions degenerate into highly-coupled conflict resolution problems. While Multi-Agent Reinforcement Learning (MARL) offers a potential solution for this problem, it is typically hindered by the dimensionality curse in large-scale coordination, poor generalization across stochastic traffic distributions, and training instability in congested maritime environments. To address these systemic bottlenecks, this paper proposes GAT-WMF-SAC, a cooperative control framework for bottleneck scenarios. By embedding Graph Attention Networks (GAT) and Weighted Mean-Field (WMF) into a soft actor-critic (SAC) architecture, it transforms myopic conflict resolution into system-level operational stability through advanced situational awareness. Our approach introduces three key contributions: 1) a WMF Q-function that adaptively estimates interaction weights to mitigate dimensionality explosion in large-scale port traffic conditions; 2) a GAT-driven feature extractor that uniformly processes dynamic maritime traffic topologies with fluctuating participant densities, enhanced by a dual-coordinate transformation and curiosity-driven rewards to encourage spontaneous deconfliction and prevent deadlocks; and 3) a centralized training paradigm with a decentralized execution framework that fuses macroscopic WMF spatial interaction features with topological traffic observations to ensure stable value estimation within complex maritime traffic patterns. In benchmark multi-vehicle coordination scenarios, GAT-WMF-SAC achieved a 100% success rate, improved trajectory smoothness by 18.2%, and reduced curvature variance by 40.7% over the baseline, with all other metrics being optimal or near-optimal, demonstrating its significant superiority in managing complex maritime traffic.