A Cooperative Navigation Framework for Multi-USV in Maritime Intelligent Transportation Systems Based on LTD–MATD3

In the field of maritime intelligent transportation, cooperative multi-USV tasks in traffic dense near-shore waters require balancing navigation progress and safe collision avoidance under short range interactions. To address this problem, this paper proposes a task-decomposed framework for multi-USV cooperative navigation, whose core decision module is a Task-Decomposed Multi-Agent Twin Delayed Deep Deterministic Policy Gradient method with long short-term memory (LSTM), termed LTD–MATD3. The proposed method constructs a unified task-decomposed mechanism to coordinate navigation progress and collision avoidance consistent with the International Regulations for Preventing Collisions at Sea (COLREGs). Specifically, it models a progress branch and a safety branch in value learning, while generating a unified continuous action through feature fusion, thereby reducing interference between progress and safety objectives. In addition, temporal sequence encoding and prioritized experience replay (PER) are incorporated to improve decision stability and sample efficiency under partial observability. Simulation results under different fleet sizes and obstacle densities show that the proposed method achieves faster convergence and higher cumulative returns than representative multi-agent reinforcement learning algorithms. Experiments in electronic nautical chart environments further verify its stability and generalization.