One camera, no expensive sensors — deploying robots just got cheaper

Getting robots to navigate used to require expensive hardware — LiDAR units, depth cameras, sensor arrays that often cost more than the robot itself. Mistral AI's new model, Robostral Navigate, changes that: a software model now replaces the sensor stack. A robot running this model can find its way through offices, warehouses, and outdoor spaces using just one ordinary camera — the kind built into any phone.

The model is small by today's standards: 8 billion parameters, similar in size to models you can run on a laptop. But its results are strong. On a standard test for navigating unfamiliar buildings, it scored 76.6% — nearly 10 points ahead of every other single-camera system, and even better than systems that use expensive depth sensors and multiple cameras. The whole thing was trained entirely in simulation: 400,000 practice runs across 6,000 virtual scenes, with no physical robots needed.

Airbus and BMW have already signed on as customers. Those are not companies that try new things early — they take years to approve anything entering their supply chain. Deals with those two signal real industrial adoption, and it matters. Mistral, a French company that raised $830 million earlier this year, is making a clear bet: the market for AI that controls physical machines is as large as the market for AI that handles text and images.

For any business running, building, or planning a robot deployment — in logistics, delivery, manufacturing, or hospitality — the message is simple: deploying capable robots just got cheaper. You no longer need a full sensor suite to get reliable autonomous navigation. One camera and the right software is now enough, which makes it realistic to put robots in many more locations at much lower cost.