MIT-SPARK/VGGT-SLAM
VGGT-SLAM is a real-time dense RGB SLAM system that uses deep-learning-based visual models (VGGT, SAM 3, Perception Encoder) for scene reconstruction.
Velocity · 7d
+2.6
★ / day
Trend
→steady
star history
The system performs simultaneous localization and mapping (SLAM) using feed-forward deep-learning models optimized on the SL(4) manifold. It leverages a visual general Gaussian tracking backbone (VGGT) and segmentation models (SAM 3) to produce dense 3D scene reconstructions from RGB input in real time.