Trained on voyage data from a ship in Japan’s crowded Seto Inland Sea, a diffusion-based navigator balanced safety, efficiency, and maritime law on its own — and started keeping right in a traffic lane without being told to.

Featured image credit: Tom Fisk/Pexels

Self-driving cars get most of the attention, but in some ways teaching a machine to steer a ship is the harder problem. In busy waterways like Japan’s Seto Inland Sea, vessels thread through dense traffic, narrow channels, and hundreds of islands, under conditions that shift constantly, while obeying international and local navigation rules.² Experienced mariners handle all of it at once, often making judgment calls that are difficult to reduce to equations.

A research group led by Assistant Professor Takefumi Higaki of Osaka Metropolitan University’s Graduate School of Engineering decided not to try. Instead of writing down rules and objectives for an AI to optimize, they trained one to learn from what real captains actually did. The resulting system, described in the journal Ocean Engineering, handled situations that confused conventional AI navigators — and picked up a local custom that was never programmed into it.^1,2

Learning from the Fukae-Maru

The researchers trained their model on maneuvers performed by the Fukae-Maru, a training vessel operated by Kobe University that regularly sails the congested Seto Inland Sea.² The crew’s decisions in those waters carried exactly the kind of ambiguity and human judgment that is hard to capture with fixed rules.

Rather than a conventional AI that predicts the single best next action, the team used a diffusion model, the same broad family of AI techniques behind many image generators. A diffusion approach works with the full range of actions experienced humans might plausibly take and generates an entire trajectory at once, rather than choosing one move at a time.²

Outperforming the competition

The researchers pitted their system against two leading navigation AIs built with conventional imitation-learning methods. In a realistic simulation, the diffusion-based model managed situations that tripped up the other two. It could deal simultaneously with any number of other ships, coastlines, narrow waterways, and speed control. In ship encounter tests, it consistently complied with the international regulations for preventing collisions at sea, keeping a safe distance from other vessels.²

“The most distinctive feature of this study is that we did not explicitly balance multiple objectives such as collision avoidance, geographical constraints, navigation efficiency, and compliance with maritime traffic rules; however, the AI achieved them,” Higaki said.²

Keeping right in the Akashi Strait

The most surprising result emerged as training progressed. The AI began exhibiting behavior the researchers never coded, including local navigation customs. When passing through the Akashi Kaikyo Traffic Route — the busy channel beneath the Akashi Kaikyo Bridge between Kobe and Awaji Island — vessels are expected to keep to the right within designated traffic lanes. Although no one programmed that convention into the model, it consistently steered the ship into the correct lane.²

That kind of emergent behavior is exactly what the data-driven approach was meant to capture. Human captains carry a great deal of tacit knowledge about how specific waters are sailed, much of it never formally written into algorithms. A model that absorbs those habits from real voyage data may be better prepared for the messy reality of crowded seaways than one built on idealized rules.

“Our approach enables the AI to autonomously learn sophisticated ship-handling skills directly from real-world operational data rather than having researchers manually define what constitutes correct behavior,” Higaki said.²

A response to a shrinking workforce

The researchers expect autonomous navigation systems to become more common as more real-world vessel operation data becomes available, much as AI has spread into cars in recent years. They hope their work will improve navigational safety and help address worsening labor shortages in the maritime industry, especially in Japan.²

There are, of course, open questions. A system that learns from human behavior also inherits human habits, good and bad, and its performance depends on the breadth of the data it sees. A model trained on one vessel in one sea will need to be tested across many ships, routes, and weather conditions before anyone hands it the helm. But the study suggests a different philosophy for autonomous shipping: rather than trying to out-calculate experienced captains, build machines that learn the way apprentices do — by watching the people who already know the water.

Higaki, Nobukazu Wakabayashi, and Hirotada Hashimoto received financial support from the Japan Society for the Promotion of Science for the work.¹

---

Notes

1. Diffusion route planner: Data-driven modeling of human ship navigation that implicitly balances safety, efficiency, and rule compliance under complex geographical constraints. (2026). Ocean Engineering. https://doi.org/10.1016/j.oceaneng.2026.125656
2. Osaka Metropolitan University. (2026, July 22). Experienced captains vs. conventional AI: Setting a new course for autonomous ship navigation [Press release]. EurekAlert! https://www.eurekalert.org/news-releases/1136883