EgoPro

1,000 hours of synchronized head-and-wrist human activity for fine-grained manipulation research.

Part of EgoSuite-Open10K · 3 configs · 7 scene families · dual-view capture + progressive supervision

Collection · EgoStandard · Open Demo · 中文 · EgoSuite

Public landing card · manual gated access. This repository is now publicly discoverable, while access requests use manual review. The Dataset Viewer remains disabled until the audited Parquet manifests, media shards, approved license, and final citation are uploaded. This public card does not itself grant rights to unpublished training data.

EgoSuite-Open10K at a glance

EgoSuite-Open10K is Lightwheel's 10,000-hour egocentric human-data release for embodied AI and world-model research. It is organized as one Hugging Face Collection and three repositories, with a shared data contract across all six formal SKUs.

Repository	Role	Configs	Formal hours	Access at release
EgoStandard	Head-view series	EgoStandard / EgoStand-motion / EgoFull	9,000	Manual gated
EgoPro	Head + wrist series	EgoPro / EgoPro-motion / EgoProMax	1,000	Manual gated
EgoDemo	Open duplicate trial pack	EgoDemo	20–50, excluded from total	Public, ungated

The official 10,000-hour total is 9,000 h + 1,000 h. EgoDemo duplicates selected source clips and is never added to that total.

Three configs in this repository

Config	Format	Hours	Share of this repo
EgoPro	Head + wrist video + 3D hand pose	750	75%
EgoPro-motion	Head + wrist video + 3D hand pose + 3D body pose	200	20%
EgoProMax	Head + wrist video + 3D hand pose + 3D body pose + V7 semantics	50	5%
	Repository total	1,000	100%

Dual-view alignment

Every Pro-series row binds the synchronized viewpoints and supervision layers to one stable clip_id. Motion and semantic tiers add references to the same clip contract; they do not create extra copies of the head or wrist videos.

clip_id
├── head_video_ref
├── wrist_video_ref
├── synchronization_ref
├── hand_pose_3d_ref
├── body_pose_3d_ref        # motion and max tiers
└── semantics_v7_ref        # max tier

Load a config

from datasets import load_dataset

pro = load_dataset(
    "LightwheelAI/EgoPro",
    "EgoPro",
    split="train",
    streaming=True,
)

motion = load_dataset(
    "LightwheelAI/EgoPro",
    "EgoPro-motion",
    split="train",
    streaming=True,
)

pro_max = load_dataset(
    "LightwheelAI/EgoPro",
    "EgoProMax",
    split="train",
    streaming=True,
)

The lightweight Parquet manifests are the config entry points. Large media and annotation objects remain versioned once in canonical paths and are resolved through the reference columns.

Common manifest contract

Field	Type	Meaning
clip_id	string	Globally unique, anonymous clip key
config_name	string	Exact HF config / SKU name
scene_family	string	One of seven top-level scene families
task_name	string	Normalized task name
duration_s	float32	Audited usable duration in seconds
head_video_ref	string	Canonical head-view media reference
wrist_video_ref	string	Canonical synchronized wrist-view reference
synchronization_ref	string	Timing/calibration metadata shared by both views
hand_pose_3d_ref	string	Frame-aligned 3D hand-pose reference
body_pose_3d_ref	string, nullable	Frame-aligned 3D body-pose reference
semantics_v7_ref	string, nullable	Frame-aligned V7 semantic reference
release_revision	string	Immutable release revision
sha256	string	Integrity checksum for the row's primary media object

Config-specific required and nullable fields are defined in metadata/schema.json.

Repository layout

EgoPro/
├── README.md
├── README_zh.md
├── assets/
├── manifests/
│   ├── EgoPro/part-*.parquet
│   ├── EgoPro-motion/part-*.parquet
│   └── EgoProMax/part-*.parquet
├── media/head/<shard>.tar
├── media/wrist/<shard>.tar
├── metadata/synchronization/<shard>.tar
├── annotations/hand_pose_3d/<shard>.tar
├── annotations/body_pose_3d/<shard>.tar
├── annotations/semantics_v7/<shard>.tar
└── metadata/
    ├── sku_catalog.json
    ├── schema.json
    ├── statistics.json
    └── checksums.sha256

Scene coverage

The release spans seven top-level families: Home, Hospitality, Retail, Sports, Logistics, Office, and Industry. Final per-scene hours must be generated from the audited release manifest rather than estimated in the Dataset Card.

Access and responsible use

This is a formal-release repository. This public repository uses manual gated access. Access approval does not override the dataset license or use restrictions published with the release.

Users must not attempt to identify participants, reconstruct sensitive locations, or use the data for surveillance, profiling, or harmful applications. The release documentation must stay version-aligned with participant authorization, anonymization, privacy review, quality-control records, and the audited manifest.

License, citation, and contact

The approved dataset license and citation author list are still pending. Public page visibility and manual access review do not grant rights to unpublished training data.

For product information, visit Lightwheel EgoSuite. For release or collaboration inquiries, use the official Lightwheel contact form.

@dataset{lightwheel_egosuite_open10k_egopro_2026,
  author    = {{Lightwheel}},
  title     = {EgoSuite-Open10K: EgoPro},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/xuejf/EgoPro}
}

Citation metadata is a release-candidate template until the approved author list and publication record are locked.