Robostral Navigate
Robostral Navigate, an 8B model, steers robots with one RGB camera and language instructions and scores 76.6% on unseen R2R-CE environments.
Mistral's first embodied model, trained entirely in simulation (2.4M trajectories over 350,000 scenes) with supervised prefix-caching then online RL (CISPO). It predicts image coordinates to move toward, so one model serves wheeled, legged and flying robots. Beats multi-sensor systems by 4.5 points.
- Date
- Wednesday, 8 July 2026
- Lab
- Mistral AI
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| R2R-CE success rate (unseen) | 76.6% 79.4% on seen; +9.7 over best single-camera, +4.5 over multi-sensor | company |
Licensing and availability not stated in the post ('talk with our team'); technical report cited as arXiv 2607.20785 (not opened). Bloomberg link is from search results, not opened.
Sources
- mistral.ai/news/robostral-navigate/
- www.bloomberg.com/news/articles/2026-07-08/mistral-ai-releases-robotics-model-to-support-p
This record was checked against its sources on 6 October 2026. How we check