Datasetpublic

Hubdata/egocentric-stereo-rgbd

Hub Egocentric: Stereo RGB-D 9 egocentric stereo RGB-D clips with dense metric depth, IMU, 6-DoF VIO pose, and calibration. Each clip carries a self-contained LeRobot v3.0 dataset (loads on lerobot >= 0.6.0) plus side-by-side stereo, mono, a colorized depth preview, and a Foxglove MCAP recording. Part of the Hub Egocentric Human Demonstrations Sample Set collection. Captured on StereoLabs ZED X Mini. Egocentric, human-demonstration data (passive; no robot action stream). July… See the full description on the dataset page: https://huggingface.co/datasets/Hubdata/egocentric-stereo-rgbd.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

1likes982downloads

Dataset Card

Hub Egocentric: Stereo RGB-D

9 egocentric stereo RGB-D clips with dense metric depth, IMU, 6-DoF VIO pose, and calibration. Each clip carries a self-contained LeRobot v3.0 dataset (loads on lerobot >= 0.6.0) plus side-by-side stereo, mono, a colorized depth preview, and a Foxglove MCAP recording.

Part of the Hub Egocentric Human Demonstrations Sample Set collection. Captured on StereoLabs ZED X Mini. Egocentric, human-demonstration data (passive; no robot action stream). July 2026.

Dataset structure

Each clip is a top-level folder named #NN_... holding its media, a Foxglove Data.mcap, the sensor sidecars, and a per-clip Metadata.yaml. A machine-readable catalog.parquet (rendered below) indexes every clip.

Catalog

clip_id	task	device	environment	duration_s	fps	resolution
#01StereoRestaurantChopVegetables	Chop Vegetables	StereoLabs ZED X Mini	Restaurant	59.84	30.08	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#02StereoRestaurantArrangeTable	Arrange Table	StereoLabs ZED X Mini	Restaurant	163.37	29.38	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#03StereoRestaurantArrangeGlasses	Arrange Glasses	StereoLabs ZED X Mini	Restaurant	66.4	28.91	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#04StereoRestaurantWashDishes	Wash Dishes	StereoLabs ZED X Mini	Restaurant	57.08	29.43	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#05StereoMeetingRoomArrange_Chairs	Arrange Chairs	StereoLabs ZED X Mini	Meeting Room	60.45	60.0	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#06StereoOffice_Photocopy	Photocopy	StereoLabs ZED X Mini	Office	74.98	60.0	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#07StereoOpenspaceRackThe_Ball	Rack the Ball	StereoLabs ZED X Mini	Open Space	66.12	60.0	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#08StereoOutdoorSeatingArea_Sweep	Sweeping	StereoLabs ZED X Mini	Outdoor	100.77	29.77	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)
#09StereoPantryWashTea_Set	Wash Tea Set	StereoLabs ZED X Mini	Pantry	74.42	60.0	1920x1200 per eye (Stereo.mp4 is side-by-side 3840x1200)

Quality note: clip #03 (Arrange Glasses) has greater capture-motion/depth temporal variation than clip #01 (Chop Vegetables). Sequential review found no alternating-frame, freeze, blank, or decode defect. The media is retained unchanged as a temporal-variation evaluation case.

Loading (LeRobot v3.0, per clip)

Each clip folder contains a self-contained LeRobot v3.0 dataset under LeRobot_v3.0/, loadable on lerobot >= 0.6.0 with native metric-depth decode:

python

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

clip = "#01_Stereo_Restaurant_Chop_Vegetables"
local = snapshot_download(
    "Hubdata/egocentric-stereo-rgbd",
    repo_type="dataset",
    allow_patterns=[f"{clip}/LeRobot_v3.0/**"],
)
ds = LeRobotDataset(
    "Hubdata/egocentric-stereo-rgbd",
    root=Path(local) / clip / "LeRobot_v3.0",
    revision="main",
    depth_output_unit="mm",  # use "m" for metres; frames and stats rescale together
)
frame = ds[100]  # observation.images.left/right/depth, observation.state (IMU), vio_pose

Each clip keeps its native 29, 30, or 60 FPS timing. The per-clip roots are intentional because LeRobot declares one global FPS per dataset. The 12-bit depth video decodes to one-channel metric depth via the loader; it is not meant to preview in the browser. See docs/ for the full field contract.

Documentation

Full delivery docs are in docs/: DELIVERY_README.md (the complete delivery contract, including the LeRobot v3.0 load path and per-feature field reference) and DATA_CARD.md (content, hardware, depth provenance, accuracy, and limitations).

Terms

Provided by Hub Data, Inc. (hub.xyz). See LICENSE. Contact Hub Data, Inc. (hub.xyz) for data-use terms.

Citation

bibtex

@misc{hub_egocentric_2026,
  title  = {Hub Egocentric Human Demonstrations},
  author = {Hub Data, Inc.},
  howpublished = {\url{https://hub.xyz}},
  year   = {2026},
  note   = {Egocentric human-demonstration sample delivery}
}

Work with Hub: https://hub.xyz/