TraceExtract open-sourced: data engine for µ0 world model trained on video with zero action labels
RexDouglass · x · 2026-10-09
A University of Maryland and Seoul National University team released µ0 (Mew-Zero), a scalable 3D interaction-trace world model (CoRL 2026), and open-sourced its data engine TraceExtract (Apache-2.0). µ0 trains purely on video with zero action labels yet rivals VLAs trained on massive action-labeled datasets, learning a "physical language" for robots.
TraceExtract automatically extracts from raw video:
- Depth plus camera intrinsics/extrinsics
- DINOv2-based semantic keypoints tracking what actually moves
- Global-local 3D reconstruction for long videos without OOM
- Event-centric VLM task descriptions
More from Embodied
- Atomic Machines claims a programmable factory for manufacturing microscopic machines — jimmystar889 · 2026-10-09
- RoboStrategy to Ring Nasdaq Closing Bell with Dyna Robotics Humanoid at Founders Forum — Rewkang · 2026-10-09
- New Paper: Native Action-Prior Learning from Videos for World Action Models — _akhaliq · 2026-10-09
- Shenzhen Hotel Robot Delivers Surprise Birthday Cake, But "No Human Involved" Isn't Quite True — EleanorOlcott · 2026-10-09
- Open-Source Models Power Home Robot Tidying for Toddlers in Weeks, Not Years — chris_j_paxton · 2026-10-09
- Sony's quadruped walks on cracked pavement with passive-stability spherical feet — BLUECOW009 · 2026-10-09