Robostral Navigate aims to guide robots using one RGB camera

Mistral introduced Robostral Navigate on 8 July, describing an eight-billion-parameter model that uses an RGB camera and a language instruction to guide a robot. The system predicts a target point in the camera view and a desired orientation, falling back to local movement instructions when a target is outside the image.

The model was trained in simulation, and Mistral reported a 76.6% success rate on the unseen validation portion of R2R-CE. That benchmark result does not mean a robot can navigate every unfamiliar building reliably. The announcement describes use across different robot bodies and camera setups, with intended applications in logistics and hospitality. Hardware integration, obstacle handling and the environment’s operating rules remain material parts of turning a navigation model into a deployed robot.