Mistral AI Unveils "Robostral Navigate" for Robots — Autonomous Navigation Using Just a Single Camera
On July 8, 2026, Mistral AI unveiled “Robostral Navigate,” an autonomous navigation model for robots. Without relying on depth sensors or LiDAR, the model lets robots move autonomously through complex environments using just a single standard RGB camera and natural-language instructions — a launch that signals the company’s full-fledged entry into physical AI.
Details
- Model scale: 8 billion parameters, trained on roughly 2.4 million movement trajectories and 350,000 scenes in simulated environments
- Sensor requirements: Operates using only a monocular RGB camera, with no depth sensors or multiple cameras required. Supports various robot forms, including wheeled, legged, and flying types
- Benchmark performance: Achieved a 76.6% success rate on R2R-CE (Room-to-Room in Continuous Environments) evaluation in unseen environments — 9.7 points ahead of the best existing monocular-camera method, and even 4.5 points ahead of methods using depth sensors or multiple cameras
- Training approach: Uses pointing-based navigation, designed to allow continuous improvement through reinforcement learning
- Development: Built in-house at Mistral AI from the ground up. Designed to be hardware-agnostic, allowing deployment across any robot fleet
How to try it
- No timeline for general availability or API release has been announced yet; Mistral AI is asking interested companies to “contact the team”
- Licensing terms have also not been specified in the official announcement