SLAM3R: Real-Time 3D Scene Reconstruction from Single Video

mhdfaran · x · 2026-07-19

Chinese researchers open-sourced SLAM3R, which can reconstruct a complete 3D scene in real time from a single smartphone video. It requires no LiDAR, depth cameras, or multi-camera setup; input is just a monocular RGB video, and output is a usable 3D world.

The key to this method is no per-frame pose optimization; instead, it directly regresses 3D geometry using a feedforward neural network, achieving 20+ FPS on a single GPU. The post also notes it outperforms traditional optimization methods on reconstruction benchmarks and does not easily drift in long walkthrough scenes.

Related event: SLAM3R: Real-Time 3D Scene Reconstruction from Monocular Video(4 posts)→

Original post →

More from Embodied

Embodied channel →