Mistral unveils Robostral Navigate, an 8B model that steers robots using just a single RGB camera

Mistral has launched Robostral Navigate, its first model purpose-built for embodied robot navigation. The 8B-parameter system takes a single RGB camera feed and a plain-language command — "walk through the corridor, enter the supply room, stop to face the second shelf" — and guides a robot through the environment autonomously.

The model posts state-of-the-art results on the R2R-CE benchmark, scoring 76.6% on validation scenes it never saw during training. That beats the best competing single-camera approach by 9.7 points and even outperforms systems that rely on depth sensors or multiple cameras by 4.5 points — all while using neither. The model works on wheeled, legged, and flying robots and generalizes across different robot sizes.

Robostral Navigate was built entirely in-house and does not depend on existing open-source vision-language models. Mistral bootstrapped it from a VLM already specialized in grounding tasks like pointing, counting, and object localization, then extended those spatial priors into full navigation. The team generated roughly 2.4 million training trajectories across 350,000 simulated scenes and used a prefix-caching strategy that compresses whole episodes into single sequences, cutting training token counts by 22× and turning multi-month training runs into multi-day ones.

After supervised training, Mistral applied online reinforcement learning with its CISPO algorithm, which lets the model recover from mistakes and explore — a move that lifted success rates by another 3.2% with no plateau in sight. For builders in logistics, delivery, manufacturing, or hospitality, Robostral Navigate offers a compact, sensor-light model that can follow a single high-level instruction and carry it out in live, obstacle-filled spaces it was never shown.