UniHand

BeingBeyond's dataset that unifies many kinds of human hand video into hand-motion annotations, for pretraining dexterous-hand VLA models.

UniHand was released in July 2025 by BeingBeyond (智在无界), together with Peking University and Renmin University of China, alongside the Being-H0 model. It unifies human hand data from 11 sources: motion-capture data (such as ARCTIC, HOI4D, and DexYCB), VR-recorded data (such as EgoDex), and RGB-only video (with hand pose estimated using the HaMeR model) — all converted into MANO parameters (a parametric hand model that represents hand shape and pose with a few dozen numbers), totaling over 1,100 hours of video and more than 150 million instruction samples. It addresses the scarcity of real dexterous-hand robot data: a model first learns “how a hand should move” from human hands, then transfers that to a robot. The January 2026 Being-H0.5 expanded it into UniHand-2.0; UniHand_Preview on Hugging Face is a subset of this pretraining data, packaged in WebDataset format.

ExampleAfter pretraining on UniHand, Being-H0 can be given an image and the instruction “unplug the AirPods charging cable” and generate a sequence of hand motions for the next few seconds for both hands.

Also called

UniHand-1.0, UniHand_Preview

Related

Being-H0 · MANO · Human Video Data · Pretraining on Human Videos · Dexterous Manipulation · EgoDex

Sources

arXiv 2507.15597: Being-H0
BeingBeyond/UniHand_Preview (Hugging Face)
Being-H0.5 项目页 (Chinese)

As of

2026-01