tl;dr: Mistral's Robostral Navigate is an 8B vision-language model for robot navigation that uses only a single RGB camera (no LiDAR or depth sensors) and achieves 76.6% success on the R2R-CE unseen benchmark, beating multi-sensor approaches by 4.5 points. It works by "pointing" to target coordinates in the camera view, was trained entirely on ~400K simulated trajectories using prefix-caching (22× token reduction), and was further improved via CISPO online RL. It generalizes across wheeled, legged, and flying robots.