A new model from Mistral suggests that robots may not need expensive sensor suites to navigate the physical world — just one ordinary camera and a sentence.
For decades, the dream of a helpful home or warehouse robot has been held back by a quiet problem: getting the machine to understand where it is and how to move through a cluttered, unpredictable room. Most approaches bolt on depth sensors, LiDAR, or clusters of cameras. A new model called Robostral Navigate takes the opposite bet — and it works with a single everyday RGB camera.
Robostral Navigate is an 8-billion-parameter model (the "8B" refers to the count of learnable numbers inside it) built specifically for embodied navigation. You feed it two things: a stream of images from one camera, and a plain-language instruction such as "Leave the lobby, walk through the corridor, enter the supply room, and stop to face the second shelf." The model then steers the robot step by step toward the goal.
The benchmark that matters here is R2R-CE (Room-to-Room in Continuous Environments), which measures how well a system follows instructions in spaces it has never seen during training. On the held-out validation set, Robostral Navigate reaches 76.6% success — and it does so without any depth sensor, outperforming systems that rely on multiple sensors working together.
Depth cameras, LiDAR, and multi-camera rigs are expensive, power-hungry, and fragile. By proving that one cheap RGB camera is enough, Robostral Navigate lowers the cost and complexity floor for physical AI dramatically. A robot built on this approach could, in principle, use the same camera a phone already has.
The model was trained entirely in-house using simulated data — virtual environments rather than real-world footage — combined with token-efficient techniques that keep the model small enough to run on modest hardware. Despite learning in simulation, it generalizes to real-world obstacles it never encountered during training, and it adapts across different robot body types.
Two ideas work together. First, a pointing-based navigation scheme lets the model indicate where to go next in the image. Second, reinforcement learning allows the system to refine its behavior through trial and error, nudging it toward strategies that actually reach the destination. The combination is aimed at a longer-term goal Mistral calls "unified embodied AI" — one model family that can both reason and act in the physical world.
Robust navigation is only one ingredient in a useful robot, but it is a foundational one. If a single camera and a sentence are enough to get a machine from point A to point B in an unfamiliar building, the path toward affordable, deployable physical AI suddenly looks a lot shorter.