ArtHOI accepted to TPAMI: turns AI-generated video into supervised-free 4D articulated human-object interaction data
liuziwei7 · x · 2026-10-25
ArtHOI, a joint project from HUST, NTU, BAAI, Tsinghua and ZJU, has been accepted to IEEE TPAMI. It converts diffusion-generated monocular 2D videos into physically grounded 4D articulated human-object interaction (HOI) data — with no 3D motion supervision required.
Key points:
- Formulates articulated HOI synthesis as 4D reconstruction from monocular video priors, treating generated video as supervision for an inverse rendering problem, so contact, articulation and temporal coherence are respected naturally
- Stage 1: flow-based part segmentation maps optical-flow cues and SAM masks onto 3D Gaussian parts, using quasi-static binding to separate moving and static regions
- Stage 2: decoupled reconstruction recovers articulated object geometry first, then synthesizes the full interaction scene
The system can synthesize realistic interactions with articulated objects like doors, cabinets, fridges and microwaves while maintaining proper contact and kinematic constraints. Project page, arXiv paper and code are public.
More from Embodied
- Kindle hack adds AI RSS, one-sentence e-book delivery, QR-code API key setup — vista8 · 2026-10-25
- Three Armenian teens win WRO European Open with a robot that helps visually impaired classmate read — RevLebaredian · 2026-10-25
- Apple's foldable iPhone hinge was AI-designed, but early units show screen crease issues — flngr · 2026-10-25
- Maker replicates Hugging Face MicroDuck robot for ~$400 with custom servo control — TinfoilTricorn · 2026-10-25
- Nvidia open-sources Flora, cuts robot video labeling cost 125x vs GPT-6 Astra — The Decoder · 2026-10-25
- China makes embodied AI a 15th Five-Year Plan priority, pushes humanoid mass production — ingliguori · 2026-10-25