Abstract
Reinforcement learning (RL) has shown strong potential for autonomous maritime navigation, yet current approaches remain vessel-specific: policies trained for one ship often fail on others due to differences in dynamics characteristics. This lack of generalization requires retraining for each vessel, leading to high computational cost, energy consumption, and limited scalability. This project proposes a meta-reinforcement learning (Meta-RL) framework that enables maritime navigation across multiple vessels. A multi-ship simulation database will be generated to capture physically consistent dynamic variability. A neural network will learn compact latent ship embeddings that encode intrinsic dynamic behavior for each ship. These embeddings will condition a Meta-RL controller trained over a distribution of ship models, enabling a single policy to adapt its control strategy to varying dynamics. The resulting framework will address both path-following and collision-avoidance tasks, ensuring compliance with COLREG regulations while maintaining robustness across heterogeneous vessel characteristics. The framework will be evaluated under zero-shot and few-shot transfer. By reducing retraining requirements, the project aims to enable scalable, energy-efficient deployment of AI-based maritime autonomy.
Researcher(s)
Research team(s)
Funding
Project type(s)