Mistral AI introduced Robostral Navigate, an 8 billion parameter model for autonomous robot navigation.
The system uses only one standard RGB camera to move robots through offices, homes, and outdoors. It relies on pointing instructions combined with reinforcement learning for continuous improvement.
Robostral Navigate achieved 76.6% success on the R2R-CE benchmark with unseen data. This result beats other single-camera approaches by nearly 10 percentage points. The model handles tasks like moving through corridors and facing specific shelves.
Training used approximately 2.4 million trajectories from 350,000 simulated scenes. An efficient prefix-caching algorithm reduced training tokens by 22 times compared to standard methods. Online reinforcement learning improved the success rate by an additional 3.2 percent.
