AI Learns Ship Navigation by Mimicking Human Captains
Researchers have developed a novel diffusion-based AI that learns ship navigation by analyzing experienced human captains' maneuvers, outperforming traditional AI models in complex scenarios.
Navigating a ship presents significantly more challenges than driving a car, especially in busy maritime environments like Japan's Seto Inland Sea. Vessels must contend with dense traffic, narrow channels, numerous islands, and constantly changing conditions, all while adhering to strict navigation rules.
A research team at Osaka Metropolitan University, led by Assistant Professor Takefumi Higaki, has pioneered a new approach to autonomous ship navigation. Instead of relying on pre-defined objectives and rules, their AI model was trained using the actual maneuvers of the training vessel Fukae-Maru. This "diffusion AI" focuses on generating an entire trajectory based on a range of actions observed from experienced human navigators, rather than attempting to predict a single best action like conventional AI.
This method proved effective in handling the ambiguities and human judgment often required in complex situations that are difficult to quantify mathematically. In evaluations against two leading conventional imitation-learning AIs, the diffusion AI demonstrated superior performance. It successfully managed arbitrary numbers of ships, coastlines, narrow waterways, and speed control within a realistic simulation environment.
During ship encounter tests, the AI consistently followed international collision-avoidance regulations, maintaining safe distances from other vessels. Notably, the AI began exhibiting unexpected but beneficial behaviors, such as adhering to local navigation customs like keeping right within designated traffic lanes when passing through the Akashi Kaikyo Traffic Route, despite this not being explicitly programmed. Higaki highlighted that the AI implicitly balanced multiple objectives like safety, efficiency, and rule compliance without explicit programming, learning sophisticated skills directly from real-world operational data.
This development in diffusion AI for ship navigation is significant as it moves beyond rigid rule-based systems to learn nuanced, human-like decision-making. By implicitly balancing safety and efficiency from real-world data, it offers a path toward more robust autonomous systems capable of handling complex, unpredictable environments, a crucial step for maritime safety and addressing labor shortages.
Edited by the news editor with AI from the original report — please refer to the original source.