Why iceboats?
Put simply, because we think it is something AI can probably get good at, maybe even beat people at. That is a guess for now, and finding out is part of the point.
Our first goal
This project started from a simple curiosity: could a model beat human players at iceboat racing? Our first goal is still that: train a model specialised in iceboat racing that generalises across most tracks. In other words, a driver that can be dropped onto a track it has never seen and still race it well.
Evaluation, preparedness and useful bots
Our models are nowhere near that good yet, and we want to prepare early. So alongside building a driver, we want to get better at measuring one. That means benchmarking how well agents race, and how well they hold up on tracks and conditions they have not seen.
We are also looking at the other side of that. If models do get that good, server owners will want to know who is driving, and keep their servers fair and stable. We are only starting to look into detection methods, and we expect to get plenty wrong along the way.
We also think good racing bots can be useful: NPC racers to fill a grid, coaches that help players improve, and ghosts to chase in time trials. We want to prepare for the risks and build the helpful versions at the same time.
Where we are
Early. This site will grow as we have results, demos and write-ups to share. See Research for how we plan to get there, and Readiness for how we are preparing.
