What Is Being-H0?

Being-H0 is introduced in Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos (arXiv 2507.15597) by Hao Luo and co-authors, submitted 21 July 2025. The abstract starts from a data bottleneck: existing VLAs rely on synthetic data with sim-to-real gaps or teleoperated demonstrations lacking scale and diversity, so the authors propose human hands as a "foundation manipulator". The paper through its method section, project page, README and a Hugging Face page were read; results tables were not.

Being-H0 is third-party research. Firsthand is not affiliated with its authors and makes no claim that any named model was trained on, used with or is compatible with Firsthand data. Firsthand captures first-person human video, fast, on demand and to a written spec, not robot data: 3D hand pose at 21 keypoints per hand with a per-joint visibility flag, wrist trajectories and action segments. The only language field is one task instruction per episode.

Per the paper, Being-H0 builds on the InternVL3 architecture. Pretraining teaches it to predict hand motion from an image and a language instruction. Text, vision and motion tokens share one sequence and attention. Post-training adds learnable action queries that predict an action chunk through MLP projection.

The paper describes each hand frame with the MANO model: joint angles, wrist rotation, translation and shape. It chose MANO-D162, a 6D-rotation feature that adds 21×3 joint positions. A grouped residual quantization tokenizer turns motion into discrete tokens, with separate wrist and finger tokenizers. Loss weights are Not stated in the sources read.

The paper contrasts Being-H0 with VLAs that rely on synthetic or teleoperated data, and with implicit approaches to human video such as contrastive learning, masked autoencoders and latent actions. EgoVLA, covered in its own entry, also pretrains on human video; whether either paper compares with the other is Not stated in the sources read.

The paper says human motions cannot transfer directly to robots because of morphological differences, kinematic mismatch and physical constraints, and that force and friction are missing from video data. This version does not model interactive objects such as 6D pose. The README lists training code and a simulation benchmark as planned.

Being-H0 is a model, not a data source. When a program needs first-person human video of particular tasks, environments or people, write them into a spec and confirm it on a Spec Pilot, a small fast first batch. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where it is thin. Billing is per validated hour; delivery is RLDS, LeRobot, WebDataset or HDF5.

What is Being-H0?

A dexterous vision-language-action model from a July 2025 paper, pretrained on human videos by explicitly modeling hand motion, then adapted to robots by post-training.

What data was Being-H0 trained on?

UniHand, curated from motion capture, VR and RGB-only videos. The abstract says millions of instructional instances, the introduction over 150M samples, and the project page names UniHand-2.5M. This page does not reconcile the counts.

How does Being-H0 represent hand motion?

Per the paper, with MANO-based features (MANO-D162) quantized by a part-level tokenizer with separate wrist and finger tokens.

Is Being-H0 open source?

The README lists an MIT licence, inference code, and checkpoints and a post-training dataset on Hugging Face. Training code is planned, and MANO files fall under the MANO licence. Read each licence yourself; this is not legal advice.

What robot results does the paper report?

The abstract reports expected gains in real-world manipulation as physical instruction tuning is applied. Numbers are Not stated in the sources read.

Does Firsthand use Being-H0, and can I request a custom dataset?

Firsthand is not affiliated with the Being-H0 authors and makes no claim of use, training or compatibility. You can request a custom egocentric dataset: write a spec, confirm it on a Spec Pilot, pay per validated hour.