Mistral's Robostral Navigate: A State of the Art Robotics Navigation Model

What happened

Mistral on July 8 released Robostral Navigate — a specialized 8B navigation model trained entirely on simulated data, requiring no LiDAR or depth sensors.

Context and impact

Mistral's move into robotics signals expansion beyond language. The model generalizes across wheeled, legged, and flying robots. On R2R-CE it outperforms both single-camera and multi-sensor competitors.

Details

• Architecture: 8B, trained on simulations (~400K trajectories, 6,000 scenes)
• R2R-CE: 79.4% (seen) / 76.6% (unseen)
• Beats single-camera by 9.7pts; multi-sensor by 4.5pts
• Online RL improved performance by 3.2%
• Prefix-caching: 22x reduction in training tokens
• Supports wheeled, legged, and aerial robots