LingBot-Map maps 3D scenes from one camera

This is one of the more substantive open-source 3D vision releases in a while: it pushes past offline reconstruction toward streaming, memory-backed scene understanding that can actually sit in a robotics or AR pipeline.

• –The core pitch is not just depth estimation; it combines coordinate grounding, dense geometry, and long-range drift correction in a single streaming framework
• –The repo claims ~20 FPS at 518×378 and stability over 10,000+ frames, which is the detail that matters for real deployments
• –The architecture choices, especially anchor context, pose-reference windows, and trajectory memory, point at a model designed around temporal consistency rather than one-frame accuracy
• –Apache-2.0 code plus model checkpoints makes this more useful than a paper-only release for teams building embodied AI or spatial perception systems
• –The likely ceiling is still bounded by compute and camera quality, but the direction is clear: 3D scene mapping is moving from optimization-heavy SLAM toward foundation-model style inference