World-Model-Grounded LLM Planning for AUV and ASV Navigation Near Offshore Wind Farms

Imagine a robot that can understand a simple sentence like "go inspect that wind turbine" and then figure out the steps to get there. For years, scientists have been teaching artificial intelligence to do exactly this by using large language models, the same kind of technology that powers smart assistants and chatbots. These systems are excellent at breaking down a big goal into smaller actions, like "move forward," "turn left," or "stop." However, they have a critical blind spot: they do not understand physics. They know the words for moving a boat or a submarine, but they cannot feel the weight of the water, the drag of the current, or the delay it takes for an engine to respond. If you ask a standard language model to steer a vessel around a turbine, it might suggest a turn that looks perfect on paper but fails in reality because the water pushes the boat sideways or the engine lags, causing the vessel to drift straight into the structure. This gap between what a computer thinks will happen and what actually happens in the ocean is the central problem researchers at the Norwegian Defence Research Establishment and Heriot-Watt University set out to solve.

The team developed a new way to guide autonomous robots near offshore wind farms by giving them a "world model." Think of this not as a second brain that thinks, but as a highly accurate simulator that runs in the background. In this system, the language model still acts as the commander, deciding the sequence of actions based on the mission description. But before the robot moves, the world model steps in to answer the question of "how long" and "how much." It calculates the precise duration for each movement and checks if the plan will survive the physical realities of the ocean, such as waves, currents, and the specific way the robot's thrusters work. If the plan looks like it will lead to a collision, the world model adjusts the timing or the path, ensuring the robot stays safe while still reaching its goal.

To test this approach, the researchers worked with two very different types of robots: an underwater vehicle that can move in all directions, including sideways, and a surface boat that can only move forward, backward, and turn, much like a car. They created a hybrid system where a standard physics simulator provided the basic rules of motion, and a neural network learned to correct the small, messy errors that real-world machines make. This combination allowed the system to predict the future state of the robot with extreme precision. In their simulations, the system could predict where the underwater robot would be after one minute with an error of less than a single millimeter, and where the surface boat would be with an error of less than half a meter.

The results of their tests were striking. When the robots were given missions to navigate around wind turbine towers and cables, the language model acting alone failed completely. It crashed into obstacles in every single trial because it could not account for the drift caused by water currents or the lag in the engines. In contrast, when the language model was paired with the world model, both the underwater and surface robots reached their destinations without a single collision. The system reduced the distance between where the robot stopped and its intended target by roughly 93 percent for the underwater vehicle and between 70 and 82 percent for the surface boat, compared to the unguided attempts.

A particularly clever part of the work involved the surface boat, which is often too small and cheap to carry expensive sensors like sonar or laser scanners to map the ocean floor in real time. Instead of relying on onboard sensors, the researchers taught the system to "see" the environment using satellite images, nautical charts, and weather forecasts. A vision-language model analyzed these existing maps to identify obstacles like shallow rocks or submerged structures, turning them into a digital map that the robot could use. This method proved just as effective as having a human manually draw the obstacles on a map, achieving a 96 percent accuracy rate in distinguishing navigable areas from non-navigable ones. This means that low-cost robots could soon navigate complex coastal areas without needing to carry heavy, expensive equipment.

The researchers also tested how well this system would hold up in a more realistic, high-fidelity simulation that included complex wave patterns and unpredictable currents. Even in this challenging environment, the robots remained collision-free. The system included a safety mechanism that constantly re-evaluated the plan as the robot moved. If the robot drifted off course due to a sudden wave, the system would quickly recalculate the remaining path to ensure it could still reach the goal safely. This closed-loop process, which acts like a constant check-and-balance, was crucial for success, especially in tight maneuvers where a small error could lead to a crash.

Ultimately, the study demonstrates that while artificial intelligence is powerful at understanding language and planning sequences, it needs a grounding in physical reality to be useful in the real world. By combining the high-level reasoning of a language model with a physics-aware world model, the researchers created a system that can safely guide autonomous vehicles through the complex and dynamic environment of offshore wind farms. The work suggests that the future of marine robotics lies not in making the language model smarter, but in giving it a better understanding of the physical world it operates in, allowing it to make decisions that are not just logical, but also physically possible.