Mistral's Robostral Navigate steers a robot from a single camera

Mistral has released Robostral Navigate, its first embodied model, and the notable design choice is how little it needs to see. The 8B system drives a robot's movement from a single RGB camera plus a natural-language instruction, with no LiDAR and no depth sensor. It works by predicting a point to head toward, then falling back to local displacement commands when a direct target is not clear, which keeps a language model in the loop for the high-level decision while a simpler controller handles the last step.

The numbers are competitive with heavier setups. Robostral Navigate reports 76.6 percent success on unseen validation environments and 79.4 percent on seen ones, beating the best single-camera approach by 9.7 points and, more surprisingly, edging past multi-sensor systems by 4.5 points. Doing that without depth hardware is the part worth noting, since it lowers the cost and complexity of the robots that can run it.

The training recipe is also concrete. The model learned from about 400,000 simulated trajectories across 6,000 scenes, used prefix-caching to cut token use by roughly 22 times, and added a round of online reinforcement learning with the CISPO method for a further 3.2 percent gain. A camera-only, language-conditioned navigation model from a lab known for text models is a clear signal that the frontier labs are moving into robotics, not just watching it.

Why it matters

If you work on robots, a policy that navigates from one cheap camera and plain-language instructions widens the range of hardware that can run capable navigation, without a depth-sensor budget. The sim-to-real recipe, with its trajectory counts and RL step, is specific enough to compare against your own pipeline.