Overview
- Mistral announced Robostral Navigate on Wednesday, July 8, 2026 as an 8-billion-parameter model that takes a plain-language instruction and an ordinary RGB camera feed to produce navigation actions.
- The company says the model uses a pointing-based policy that predicts image coordinates for targets and falls back to local-frame displacements when targets lie outside the camera view.
- Mistral reports Robostral Navigate was trained entirely in simulation on roughly 400,000 trajectories across 6,000 scenes using a token-saving prefix-caching method and was improved with online reinforcement learning called CISPO.
- Mistral claims 76.6% success on the R2R-CE 'validation unseen' benchmark and says the model outperforms prior single-camera and some multi-sensor systems, but those results have not yet been independently validated in real-world industrial settings.
- The release frames Robostral as a hardware-agnostic product that can run on wheeled, legged, and flying robots and fits Mistral’s shift to smaller, task-specific robotics models, though observers say real deployments and customer pilots will be the true test of sim-to-real reliability and operational safety.