Robostral Navigate uses one camera to get around

Mistral introduced Robostral Navigate, an eight-billion-parameter model that combines RGB images with natural-language instructions. The company reports navigation results using a single ordinary camera rather than depth sensors or multiple cameras, with development based on simulated data.

Context

Reducing sensor requirements could simplify deployment, but benchmark success is not a guarantee for an unfamiliar building. Lighting, moving obstacles, and the robot’s physical limits all affect the result. A navigation model needs to be assessed with the complete robot and the environments in which it will actually move.