Towards COLREGs-aware ship collision avoidance with multi-agent PPO-LSTM in maritime IoT

作者：Y. Ding, W. Meng, S. He, Weiwei Li · 发表于：Lancaster EPrints (Lancaster University) · 年份：2026 · 研究领域：Maritime Navigation and Safety、Maritime Transport Emissions and Efficiency、Underwater Vehicles and Communication Systems

Maritime Autonomous Surface Ships are expected to operate in a maritime IoT environment, where distributed sensing, V2V/AIS/VDES communication links, and electronic charts jointly support perception–decision–control loops for safe navigation in congested waters. A key challenge is to realise multi-ship collision avoidance that is consistent with the International Regulations for Preventing Collisions at Sea, while accounting for the limited manoeuvrability of large commercial vessels and the geometric constraints of ENC-derived chart-constrained narrow waterways. To address this problem, this work proposes a three-layer maritime IoT architecture in which each KVLCC2-class tanker is modelled as an IoT node, and ship states, TCPA/DCPA-based risk measures, and chart-derived environmental features are fused into a shared situational-awareness representation. On this basis, the task is formulated as a cooperative multi-agent partially observable Markov decision process, in which COLREGs encounter types, give-way/stand-on roles, and safety-domain constraints are embedded explicitly through the observation and reward design. A parameter-sharing recurrent multi-agent PPO–LSTM framework is then developed under the centralised-training-decentralised-execution paradigm, using a weakly centralised critic to handle partial observability and temporal coupling in dense multi-vessel interactions. The framework is evaluated in a unified simulation environment covering standard Imazu multi-ves...