Mistral's Robostral Navigate lets a robot see through a single camera and find its own way

Mistral AI has unveiled Robostral Navigate, its first model for autonomous robot navigation. It finds its way using video from a single ordinary RGB camera and reaches a 76.6% success rate on the R2R-CE benchmark for environments not seen during training.

What the model can do

Robostral Navigate is an 8-billion-parameter model that takes a camera frame and a text instruction and guides a robot through complex surroundings: an office, a residential or commercial building, an outdoor space. An instruction might read: "Leave the lobby, walk down the corridor, enter the storage room and stop in front of the second shelf." The model then takes the robot along the whole route on its own, without human involvement.

• 79.4% success on the R2R-CE validation set for familiar scenes and 76.6% for unseen ones
• It beats the best single-camera approach by 9.7 points and the best system using lidar or several cameras by 4.5 points
• It works on wheeled, legged and flying robots and is robust to differences in camera parameters
• The training dataset contains about 400,000 trajectories collected in 6,000 simulated scenes

How the model finds its way

Robostral Navigate does not predict a metric displacement. Instead it predicts a point in the image: the target's coordinates in the current camera frame and the required orientation on arrival. This "navigation by pointing" holds up when optics and scene scale change, unlike commands based on metric distances. When the target leaves the field of view, the model switches to displacements in the robot's local coordinates, for example: "Move 2 meters forward, 1.5 meters to the left and turn 25 degrees left."

How Robostral Navigate was trained

The model was built entirely in-house at Mistral AI and does not rely on open vision-language models. It is initialized from the company's own model specialized in pointing, counting and object localization. The key training element is a prefix-caching method: it compresses a whole episode into one sequence and lets the model train on all time steps in a single pass, cutting the number of training tokens 22-fold compared with training on individual steps. After the supervised stage, the model is further improved through online reinforcement learning with the CISPO algorithm, which lets it learn from its own mistakes and recover from failures.

"Today we are presenting

Robostral Navigate, our first model for embodied navigation," the official Mistral AI blog says.

What this means

Navigating with a single camera, without lidar or depth sensors, lowers the cost and complexity of robots for warehouses, delivery, hotels and manufacturing. Mistral AI explicitly names these industries among the first beneficiaries of the technology.

Frequently asked questions

Which robots does Robostral Navigate work on?

The model works on wheeled, legged and flying robots and generalizes across platform sizes, according to Mistral AI's description of the model.

How much data did training require?

The company collected about 400,000 trajectories in 6,000 simulated scenes. The whole data-generation pipeline was built entirely in simulation, with no real robots used during data collection.

How much more accurate is

Robostral Navigate than its competitors?

On the R2R-CE benchmark for unseen environments, the model outperforms the best single-camera approach by 9.7 points and the best system with several cameras or a depth sensor by 4.5 points, for a final 76.6% success rate.

How does the R2R-CE metric differ from ordinary navigation tests?

R2R-CE (Room-to-Room in Continuous Environments) tests how well a model follows text instructions in environments that were not part of the training set. Robostral Navigate scores 76.6% in exactly this scenario, not on familiar scenes, where the result is even higher.