Hugging Face has introduced Robostral Navigate, an 8-billion-parameter vision-language model designed to make robot navigation scalable across different platforms. The model requires only a single monocular RGB camera—the most common sensor on robots—and predicts waypoints by pointing to the next target location directly in the camera view. This approach avoids robot-specific coordinate systems, making it robust to changes in camera intrinsics and scene scale, and works on wheeled, legged, and aerial robots without recalibration.
To train the model efficiently, the team generated 2.4 million trajectories across 350,000 simulated scenes, reducing reliance on real-world data. They also introduced a prefix-caching training recipe that packs entire episodes into single training sequences, cutting training tokens by 22× and reducing training time from months to days. A tree-based attention mask prevents the model from conditioning on previous ground-truth actions, encouraging visually grounded predictions, and reinforcement learning further improves exploration and recovery capabilities.
On the Room-to-Room in Continuous Environments (R2R-CE) benchmark, Robostral Navigate achieves a 77.4% success rate, surpassing the best monocular method by 10.5 percentage points and the strongest depth- or multi-camera system by 5.3 points, despite using only a single RGB camera. On the Room-Across-Room in Continuous Environments (RxR-CE) benchmark, it reaches a 75.1% success rate, outperforming all monocular baselines.