Stanford lab rebuilt as interactive 3D web scene in a day with GPT-6 Astra and Retriever

OfirPress · x · 2026-09-10

Stanford's Linfeng Zhao used GPT-6 Astra (via a few Codex iterations) to turn a phone video of his lab's kitchen into an interactive 3D webpage: movable drawers, physics in Mujoco, clickable open/close — done in a day. The demo showcases Retriever, a compositional programming model for closed-loop robot agents: agents are temporal computation graphs where stateful Flows run on different Clocks with per-edge Sync policies, enabling asynchronous execution across backends plus a real-robot pipeline (Retriever-0). Ofir Press: imagine where this is in 2.5 years.

Original post →

More from Embodied

Embodied channel →