Mistral AI has released Robostral Navigate, its first model designed for embodied navigation using only a single RGB camera. The 8B-parameter model accepts plain-language instructions and image inputs to guide robots through complex environments like offices and outdoor settings.
- Achieves 76.6% success on the R2R-CE validation unseen split, outperforming best depth/multi-camera systems by 4.5 points.
- Uses a "pointing" method to predict target image coordinates, falling back to local metric displacements when targets are out of view.
- Trained on ~400,000 simulated trajectories across 6,000 scenes using prefix-caching, which reduces training tokens by 22×.
- Applies CISPO online reinforcement learning after supervised training, adding a 3.2% success rate improvement.
- Built from Mistral's in-house grounding VLM rather than open-source models, supporting wheeled, legged, and flying robots.
The model enables efficient navigation across diverse platforms by eliminating the need for depth sensors or LiDAR while maintaining robustness to camera intrinsics and environmental changes.