Top: RGB reconstruction and closed-set semantic segmentation.
Bottom: 3D instance memory and navigation map.
- An online, end-to-end semantic SLAM system adding LiDAR-grounding to a VGGT backbone.
- The system processes the data stream in a sliding-window fashion, creating and aligning submaps to keep VRAM usage constant.
- 2D closed-set semantic instance masks are lifted to 3D, tracked and stored in a temporal memory for change detection.