France's Mistral Launches Robot AI Model, Scores 76.6% in Navigation Test

en.Wedoany.com Reported - French AI company Mistral has released a new model called Robostral Navigate, designed to guide robots through path navigation using natural language instructions and only a single RGB camera.

Unlike most robot models that rely on depth sensors, LiDAR, or multi-camera collaboration, Robostral Navigate collects visual information using only a single RGB camera. Mistral stated that the model achieved a score of 76.6% on the R2R-CE (Room-to-Room in Continuous Environments) benchmark for instruction-following robots. Despite not using depth sensors or multiple cameras, this result is 4.5 percentage points higher than the best system using such auxiliary equipment, and 9.7 percentage points ahead of the second-best single-camera robot.

The model is designed for autonomous navigation in complex environments such as offices, residences, commercial buildings, and outdoor scenes. Mistral noted that a key feature of Robostral Navigate is its higher training efficiency, requiring significantly fewer tokens compared to other models, reducing training time from months to days.

Robotics is a mature field of AI research. Other AI model developers have already established a presence in this area, with Nvidia announcing a robot AI project in August 2025.