Mistral AI launches Robostral Navigate, 8B robot navigation model with single RGB camera
76.6% benchmark success without depth sensors or LiDAR, using only one camera.
Deep Dive
Mistral introduces Robostral Navigate, an 8B embodied navigation model that uses a single RGB camera and plain-language instructions to move robots. Achieving 76.6% success on unseen R2R-CE benchmarks, it outperforms multi-sensor systems. Trained on approximately 400,000 trajectories collected across 6,000 scenes—all generated in simulation—the model generalizes across wheeled, legged, and flying robots, unlocking applications in manufacturing, delivery, logistics, and hospitality.
Key Points
- 76.6% success rate on unseen R2R-CE benchmarks, beating multi-sensor systems by 4.5 points.
- Uses only a single RGB camera — no LiDAR, depth sensors, or multiple cameras required.
- Trained entirely in simulation on 400,000 trajectories; generalizes across wheeled, legged, and flying robots.
Why It Matters
Mistral's first embodied AI model proves that language-guided, camera-only navigation can outperform sensor-heavy approaches, lowering robot costs and unlocking scalable autonomy.