Curriculum-guided RL framework boosts autonomous ship navigation in ports

As smart ports increasingly rely on IoT-enabled autonomous vessels, ensuring reliable navigation under dense traffic and partial observability remains a critical challenge. In their arXiv paper (2608.11597), Yuqing Lin, Rangya Zhang, and Kum Fai Yuen propose a curriculum-guided reinforcement learning framework with a shared recurrent policy. The design separates centralized training—done offline as a design-time strategy—from onboard execution, meaning all navigation actions run locally on edge devices. This aligns with IoT edge intelligence paradigms, reducing latency and dependence on network connectivity.

The framework leverages curriculum learning to progressively train the policy on increasingly complex scenarios, enhancing temporal reasoning and robustness. Extensive simulations across multiple realistic port environments show that the approach outperforms standard baselines in navigation reliability, collision avoidance, and training stability. Crucially, it generalizes effectively to previously unseen high-density situations, a key requirement for real-world deployment. The authors argue this shared policy learning methodology offers a scalable, practical solution for autonomous maritime devices, potentially accelerating the adoption of self-navigating vessels in smart port operations while maintaining safety and efficiency in congested waterways.