Robostral Navigate tells a robot where to go by looking through a single ordinary camera. No laser rangefinder. No depth sensors. No several cameras at once. The robot orients itself roughly the way we do - with eyes and experience.
Fact
The facts: an 8-billion-parameter model, trained entirely in simulation, about 400,000 trajectories across 6000 scenes, never once with a real robot in a real room. According to Mistral, on the R2R-CE benchmark the model achieves 79.4% success in familiar scenes and 76.6% in unfamiliar ones, which the company says is 9.7 points above the best previous single-camera approach and 4.5 points above systems using depth sensors or multiple cameras. A caching technique cuts the tokens needed for training 22 times over, also according to Mistral. No price or license has been announced; the invitation is to talk with the team, meaning access is through partnership, not a free download.
What actually works shows up in what you choose to remove, not in how many points you beat the leaderboard by. The laser rangefinder and depth sensors often cost as much as the entire chassis - they're the expensive half of any mobile robot. If around eighty percent success holds up with one cheap camera outside the test zone, the threshold for who can build robots drops sharply. That math interests me more than the percentage.
The expensive sensor stack drops out. The robot sees with one eye, like us.
From here it's waiting. The model was trained only in simulation and isn't yet available outside partnership - the door isn't open to everyone. For now it's a promise of cheap navigation, not a machine you order tomorrow. I'll check how it behaves outside simulation when I can. Worth following for another reason too: this is the first serious move into robotics from a European player known until now only for language models.
Official primary sources
Mistral AI - Robostral NavigateThe visual is generated code-art, no third-party images.