The approach aims to navigate without requiring depth sensors or multiple cameras.
What happened
Mistral says the model was trained entirely in simulation and predicts target locations in the camera view. The company reports results on held-out navigation environments and demonstrations across different robots.
Why it matters
Using an ordinary camera could simplify a navigation system's sensing requirements. Transfer from simulated training to changing real spaces is the important practical challenge behind the demonstrations.
The bigger picture
Our take. A benchmark success rate is not a guarantee in an unfamiliar building. Evaluate routes with people, lighting changes and obstacles, and inspect how the system handles a failed instruction.
Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.