MIT Lab Kitchen Rebuilt in 3D from a 20-Second Phone Video via GPT-6 Astra

TheMoonMidas · x · 2026-09-12

Developer @frankzydou built an agentic pipeline that turns plain RGB video into 3D assets. Input: a 20-second handheld phone walkthrough of an MIT CSAIL lab kitchen — no depth sensor, no CAD, no asset library. Total time about one day with human-in-the-loop.

The loop: monocular video goes through ViPE for camera poses and metric depth (the measuring instrument, not the output), then GPT-6 enumerates and models the objects. Result: 31 modeled objects plus 6 drafts, with a live demo comparing the reconstruction against the scan. Weaknesses remain: thin and shiny objects, and the room shell needs another pass.

Original post →

More from coding & agent

coding & agent channel →