Robostral Navigate: single-camera AI navigation | Mistral AI

This post details the development of Robostral Navigate, an 8B embodied navigation model for robots. It highlights the use of a single RGB camera, achieving 76.6% success on unseen R2R-CE benchmarks by outperforming multi-sensor approaches. The model's architecture combines pointing-based navigation with local displacement commands. Key technical contributions include an in-house data generation pipeline using simulation (2.4 million trajectories), efficient supervised training via prefix-caching (reducing training tokens by 22x), and online reinforcement learning (CISPO) for performance improvement (3.2% success rate increase). The model generalizes across robot types and adapts to unseen obstacles.