PointZero: a 3D world model pre-trained without robot data hits SOTA on 3D dynamics
YunzhuLiYZ · x · 2026-09-18
PointZero introduces a simple recipe for pre-training a 3D world model without expensive robot data: learn to complete 3D point tracks, which yields transferable 3D dynamics.
- Zero-shot 3D dynamics without any robot data
- SOTA on action-conditioned 3D dynamics after post-training
- SOTA on imitation learning after post-training
The result matters for dexterous manipulation, which traditionally requires large-scale robot interaction data.
More from Embodied
- Scaled embodied foundation models enable zero-shot fully autonomous drone flight — AkshatS07 · 2026-09-18
- RLE-Bench shows robots can grow their own bodies — but coding agents lack physical reasoning — daibond_alpha · 2026-09-18
- tinydraw v2: a vector finger-drawing app on ESP32-S3 hitting ~30fps pan and 16x zoom — dejavucoder · 2026-09-18
- 40,000 US road deaths a year: the case for ending human driving ASAP — brezina · 2026-09-18
- Open-Source Dexterous Astra: Zero-Shot Pen Spinning and One-Handed Rubik's Cube in Sim — ZeYanjie · 2026-09-18
- Vacuum with 5 cameras dodges pet poop but can't recognize pee pads — yungcontent · 2026-09-18