Mistral AI to launch robotic model in physical AI push

AI company Mistral AI unveiled an 8-billion-parameter model, called Robostral Navigate, that steers robots with a single camera on Wednesday, marking the French company’s first move into physical A.

The launch is intended to challenge the sensor-heavy navigation systems used across warehouses and factories.

Mistral AI builds novel navigation system with AI

The Paris-based startup, valued at 11.7 billion euros (about $13.4 billion) in its September Series C round, has spent most of its existence competing with OpenAI on text and code. Its new model, Robostral Navigate, however, puts the company in a totally different tech category.

According to a Mistral press release, the model handles “embodied navigation,” which lets a robot move through offices, homes, commercial buildings, and outdoor spaces without external input.

However, unlike regular autonomous navigation models that lean on LiDAR, depth sensors, or several cameras working together, Robostral Navigate runs on a single RGB camera. All the model requires is a plain-language instruction just like an AI text prompt, after which it produces the movement commands to carry it out.

Mistral AI trial numbers

Mistral AI says the robotic AI model scored 76.6% on R2R-CE validation unseen, a benchmark that measures how well a robot follows instructions in environments it was not trained on. That figure beat the best previous single-camera scores by 9.7 points, according to The News International.

AI Weekly reported that the model also edged out multi-sensor systems built on LiDAR and depth by 4.5 points, with a validation-seen score of 79.4%.

If these gaps hold in real-life situations, it would be of utmost importance to those interested in industrial robots. Naturally, the general assumption in mobile robotics has been that cameras alone are too fragile, so a full sensor stack is compulsory. A build with just a single camera that matches or beats full stack systems would cut the parts bill for warehouse and facility machines.

Skepticism surrounding real-life usage

It is worth noting that these scores come from simulated situations and not real life. Mistral claims to have trained the model on about 400,000 trajectories across 6,000 scenes, using a method the company claims cut training tokens by a factor of 22, while shortening runs that took previously took months to only days.

But as AI Weekly noted, the company has not published any actual robot usage results or figures regarding on-device latency. Engineers have also questioned if a 76.6% success rate is good enough for deployment in real life.

Mistral is also reported to be in talks to raise about 3 billion euros (approximately $3.42 billion) at a valuation close to 20 billion euros (about $23 billion).

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.