Mistral's Robostral Navigate: a state of the art robotics navigation model
Introducing Robostral Navigate: 8B model achieving 76.6% on R2R-CE with just a single RGB camera. No depth sensors, LiDAR, or multiple cameras needed.
AI Summary
Robostral Navigate is an 8B AI model that enables robots to navigate complex environments using only a single RGB camera, achieving 76.6% success on the R2R-CE benchmark and outperforming multi-sensor systems. It was trained entirely in-house on simulated data using token-efficient techniques and combines pointing-based navigation with reinforcement learning to generalize across robot types and adapt to real-world obstacles. The model supports long-horizon instruction-following tasks in settings like offices, warehouses, and hospitals.