On Wednesday, February 7, MARIN gave an online technical session about an AI Sailing project that MARIN carried out in 2023. In the project an AI Sail team, consisting of a broad mix of specialists at MARIN (machine learning, time domain simulations and model testing), took up a challenge. Can a computer learn to sail an Optimist with the help of AI? Applied data scientist Fanny Rebiffé gave a presentation about the approach in this open innovation project and the results.
AI Sail aiming to train a Reinforcement Learning agent to sail an Optimist upwind
Most maritime prediction methods are based on a model-based approach: physics-based models are combined in a computational model and validated in model tests and reality. With AI Sail MARIN wanted to demonstrate the possibilities of data-driven methods, where the physics are not explicit in the model, but implicit in the data.
Fanny: “When applied in the real world, AI often suffers from a degradation of its performance learned in a simulation. With this AI sail project, we were able to address this issue. Thanks to MARIN’s complete range of facilities, we were able to create the condition for an AI to train in a simulation before entrusting it with the Optimist control.”
The AI designed for this project is a reinforcement learning (RL) agent, which learns to optimize its policy thanks to interactions with a dynamic environment. The virtual environment is a time domain simulation of the Optimist and the MARIN Offshore Basin powered by an XMF framework. The training took place in a computational cluster. “In the meantime our model test engineers had modified our Optimist with remote-controlled rudder, sheet and ballast.” In November 2023, while discovering the basin for the first time, the AI managed to transfer the strategy learned in the simulation to sail the real Optimist upwind without further training.
More information
- AI-Sail – MARIN Report No 136, June 2023
- Can a Computer Learn to Sail using AI?
- Can a computer learn to sail an Optimist with AI?
