Mistral AI has unveiled Robostral Navigate, an 8 billion parameter model that lets robots find their own way through unfamiliar spaces using nothing more than a single ordinary camera. No LiDAR, no depth sensors, no multi-camera rig, just RGB footage and a plain language instruction telling it where to go, such as directing a robot to leave a lobby and stop at a particular shelf.
The model scored 76.6 per cent on the R2R-CE unseen validation benchmark, a result Mistral says beats the best single camera system by 9.7 points and outperforms rivals using depth sensors or multiple cameras by 4.5 points. On the seen validation set it reached 79.4 per cent.
Rather than issuing metric displacement commands, Robostral Navigate works by pointing: predicting where in its current camera view the robot should head next, along with the orientation it should adopt on arrival. When the target sits outside the field of view, it falls back to local coordinate instructions instead.
Built entirely in house on around 400,000 simulated trajectories across 6,000 scenes, the model was trained using a prefix caching technique that Mistral says cut training tokens twenty two fold. A further reinforcement learning stage using an algorithm called CISPO lifted the success rate by 3.2 percentage points, with no sign yet of that improvement plateauing. Mistral frames the release as a first step toward a unified embodied AI agent for general purpose robotics.
