The great-circle route is the shortest path between two points on a sphere. For ships crossing oceans, it is the default — and it is often wrong. Ocean currents, wind, and waves create a dynamic energy landscape where the shortest path is not the most fuel-efficient. A slight detour to catch a favorable current or avoid a headwind can save more fuel than the extra distance costs.
Traditional weather routing solves this as an optimization problem: given a forecast of wind and current fields for the entire voyage, find the path that minimizes fuel consumption. The problem is the forecast. Ocean weather predictions degrade rapidly beyond 3–5 days. A two-week Atlantic crossing uses a forecast that is unreliable for most of its duration. The optimization is precise about an imprecise world, and when the forecast is wrong, the “optimal” route can be catastrophically wasteful.
PIER (arXiv:2603.17319) uses physics-informed offline reinforcement learning with a critical difference: it routes using only local observations — what the ship can see right now — rather than a global forecast. The RL agent learns a policy from historical voyage data: given current wind, current, wave height, and heading, what speed and heading adjustment minimizes fuel use? No forecast required. No global optimization. Just local decisions informed by physics.
The results: 10% mean CO2 reduction versus great-circle routing, and a ninefold reduction in catastrophic fuel waste events (from 4.8% to 0.5% of voyages). The catastrophic events are the key metric. Average improvement is modest; the elimination of worst cases is dramatic. Traditional optimization reduces the average but is vulnerable to forecast failures that produce extreme waste. Local-observation RL eliminates the dependence on forecasts entirely.
The ship that looks only at the sea in front of it navigates better than the ship that looks at the whole ocean through a cloudy window.