Live

New this week23 datasets listed, 645,000+ hSee what is listed

Home · Open data · Manipulation · Human video

Hy-Embodied-0.5-VLA-Data

2,163 hours and 250,304 episodes of bimanual demonstrations on a fingertip UMI device with motion capture, three views at 424x240; CC BY 4.0.

Published by Tencent Robotics X and Tencent Hy Team, 2026-06

Open data
Commercial use OK

CC BY 4.0

PriceFree

The publisher's licence permits commercial use and redistribution, with attribution. Hub card tag cc-by-4.0; the card's licence section says: This dataset is released under CC-BY-4.0.

Get dataset Opens huggingface.co in a new tab
From a terminal
hf download tencent/Hy-Embodied-0.5-VLA-Data --repo-type dataset --include table_000/* --local-dir ./Hy-Embodied-0.5-VLA-Data

Public, no gate. 22 self-contained tables (table_000 to table_021) of about 890 GB each; the command pulls the first. Reading the Lance tables needs the reader in the Hy-Embodied-0.5-VLA code repo or the lance library. Also mirrored on ModelScope.

Specification, as published

Hours (stated)
2,163 h
Episodes (stated)
250,304 episodes
Scale
Total Duration 2,163 hours; the open-source release contains approximately 20% of the full corpus
Task
Pick & place, Sorting & packing, Other
Modality
RGB, End-effector pose, Gripper state, Language
Capture
Egocentric
In frame
Person hands
Captured in
Asia
Embodiment
human with a custom fingertip UMI device (optical motion capture)
Frame rate
30 fps
Resolution
424x240
Formats
Lance, LeRobot v3.0, Parquet
Size
about 18.8 TB in 22 tables (20.7 TB on the Hub)
Hosted on
Hugging Face

Figures are the publisher's own. Kinetic Blocks has not measured this dataset.

About this dataset

The open slice of the corpus behind Tencent's Hy-Embodied-0.5-VLA model: bimanual manipulation demonstrated by people on a custom fingertip UMI device tracked with optical motion capture, about 20 percent of a 10,000-hour training corpus. Each step carries three camera views (head and both wrists) at 424x240 and 30 Hz, a 16-dimensional state of both end-effector poses and grippers, gripper actions and a task string in Chinese. The data ships as 22 Lance tables of about 100 hours each under a LeRobot v3.0 schema, so frames are stored as images inside the tables rather than as MP4. The resolution is low and suits policy training more than visual inspection.

Publisher's own task names: 70+ bimanual manipulation tasks, grasp the red block and put it in the box.

What Kinetic Blocks did, and did not do

Indexed and linked from the publisher. Not hosted, verified or graded by Kinetic Blocks. The publisher's terms govern. The licence shown is the one stated on the publisher's page, linked above. Nothing on this page is a claim by Kinetic Blocks.

Want the verified ones too?

Request access to the marketplace, where verified listings are checked against their files and graded A+ to F, and your money waits until the data holds up.