Mistral AI's Robostral Navigate lets robots navigate with just one camera

Mistral AI has introduced Robostral Navigate, its first embodied navigation model designed to steer robots using only a single RGB camera and plain-language instructions. The 8B-parameter model, trained entirely in simulation, achieves 76.6% success on the R2R-CE (Room-to-Room in Continuous Environments) validation unseen benchmark, beating the best single-camera approach by 9.7 points and even outperforming systems that rely on depth sensors or multiple cameras by 4.5 points. Instead of metric displacements, Robostral Navigate uses a pointing-based method: it predicts the image coordinates of the target location in the robot's current camera view, making it robust to changes in camera intrinsics and world scale. When the target is out of view, it falls back to local coordinate displacements.

Built entirely in-house and not based on existing open-source VLMs, the model was initialized from Mistral's vision-language model specialized in grounding tasks like pointing and counting. Navigation emerges as a natural extension of these capabilities. The efficient data generation pipeline produced approximately 400,000 trajectories in simulation, enabling rapid iteration. Robostral Navigate generalizes across wheeled, legged, and flying robots, and adapts to real-world obstacles unseen during training. This technology unlocks applications in manufacturing, delivery, logistics, and hospitality, making autonomous navigation more accessible without expensive sensor suites.