Mistral introduces Robostral Navigate, an 8B model for robot navigation

Mistral introduced Robostral Navigate as its first model built for embodied navigation, the company said. The 8B model takes images from an RGB camera and a plain-language instruction, then moves a robot through an environment.

Mistral says Robostral Navigate achieved 76.6% on the unseen validation split of the R2R-CE benchmark. The company says that result was 9.7 percentage points above the best single-camera approach and 4.5 points above the best system using depth or multiple cameras.

The model uses one ordinary RGB camera and no depth sensors, according to Mistral. The company describes navigation instructions such as leaving a lobby, moving through a corridor, entering a supply room and stopping while facing a specified shelf.

Mistral says Robostral Navigate was built entirely in-house using simulated data and token-efficient techniques. It combines pointing-based navigation with reinforcement learning, according to the company.

Mistral also says the model generalizes across robot types and adapts to real-world obstacles that were unseen during training. Those statements, like the benchmark comparisons and the 76.6% result, come from Mistral’s announcement; the provided evidence does not include an independent evaluation.

The announcement presents the model as a navigation system designed to work with less sensing hardware than approaches that use depth sensors, LiDAR or several cameras. The provided material does not establish how the model performs on physical robots beyond Mistral’s claim about adapting to real-world obstacles, or how it compares with systems outside the cited R2R-CE results.