Mistral enters robotics with Robostral Navigate, an 8B model that steers robots using just one camera

Mistral has unveiled Robostral Navigate, its first AI model for robot navigation. The 8B model guides robots through complex environments and, according to Mistral, needs only a single RGB camera. Still, it hits up to a 79.4 percent success rate on the R2R-CE benchmark, a standard test for navigation in unknown environments. Mistral says that beats both the best single-camera method and systems using depth sensors or multiple cameras.

The model was built entirely in-house and trained only in simulated environments, using about 400,000 recorded paths across 6,000 different virtual spaces. It works on wheeled, legged, and flying robots. Mistral hasn't shared any details on availability yet.

Mistral sees navigation as the foundation for universal robotics and plans to keep improving the model. Experiments with reinforcement learning already boosted the success rate by 3.2 percentage points, with no sign of a plateau. "We are confident that more training and more experiments will continue to push this number up," the company says.