xuejf/EgoSuite-Open1.0K

EgoSuite-Open10K

Real-world human activity, structured for embodied AI.

10,000 hours · 6 data SKUs · 7 scene families · head + wrist viewpoints · pose + semantic supervision

Overview · Six SKUs · Scene coverage · Release format · Access · 中文 · EgoSuite

Staging notice: this repository is named xuejf/EgoSuite-Open1.0K for page testing only. The product and dataset definition presented here remain EgoSuite-Open10K / 10,000 hours.

Overview

EgoSuite-Open10K is Lightwheel's planned 10,000-hour open-data release for embodied AI and world-model research. It organizes real-world egocentric human activity into six composable data SKUs, allowing teams to choose the camera viewpoints and supervision depth required by their research.

10,000 h	6	7	2	5
Release hours	Data SKUs	Scene families	Viewpoints	Modality layers

What the release is designed to provide

Dimension	Coverage	Intended research value
Real-world activity	Domestic, commercial, office, logistics, sports, and industrial workflows	Improve transfer beyond controlled laboratory demonstrations
Egocentric capture	Head-mounted RGB across all six SKUs	Preserve natural hand-object interaction and temporal context
Multi-view capture	Synchronized wrist view in the Pro series	Resolve fine-grained manipulation and hand-object occlusion
Pose supervision	3D hand pose; full-body pose in motion tiers	Support perception, action understanding, and imitation learning
Semantic supervision	Frame-aligned semantic layer in Full and ProMax	Enable structured reasoning over objects, actions, and scenes

Six Data SKUs

The Standard series scales head-view data. The Pro series adds a synchronized wrist view. Within each series, motion and full tiers progressively add body-pose and semantic supervision.

Series	SKU	Viewpoints	3D hand pose	3D body pose	Semantics	Hours
Standard	EgoStandard	Head	✓	—	—	8,400
Standard	EgoStand-motion	Head	✓	✓	—	500
Standard	EgoFull	Head	✓	✓	✓	100
Pro	EgoPro	Head + wrist	✓	—	—	750
Pro	EgoPro-motion	Head + wrist	✓	✓	—	200
Pro	EgoProMax	Head + wrist	✓	✓	✓	50
					Total	10,000

Quick SKU selection guide

• Choose EgoStandard for high-volume egocentric video and hand-pose learning.
• Choose a motion tier when full-body dynamics or action trajectories matter.
• Choose the Pro series when fine manipulation benefits from a wrist viewpoint.
• Choose EgoFull or EgoProMax when frame-aligned semantic supervision is required.

Capture and Supervision

All SKUs share a common data contract. The difference between SKUs is explicit: viewpoint coverage and annotation depth increase without changing the core organization of the release.

Scene Coverage

Scene family	Representative activity
Home	Daily routines, cleaning, cooking, and object interaction
Hospitality	Food preparation, service flows, and table operations
Retail	Picking, restocking, checkout, and customer-service workflows
Sports	Training, equipment handling, and coordinated movement
Logistics	Sorting, packing, transport, and warehouse workflows
Office	Desktop work, device use, document handling, and collaboration
Industry	Tool use, assembly, inspection, and industrial operations

Release Format

The production release is designed around versioned manifests and WebDataset shards for streaming, selective download, and integrity checking.

EgoSuite-Open10K/
├── README.md
├── README_zh.md
├── assets/
├── manifests/
│   ├── dataset_manifest.parquet
│   ├── shard_manifest.csv
│   └── checksums.sha256
├── EgoStandard/<scene>/<subject>/<shard>.tar
├── EgoStand-motion/<scene>/<subject>/<shard>.tar
├── EgoFull/<scene>/<subject>/<shard>.tar
├── EgoPro/<scene>/<subject>/<shard>.tar
├── EgoPro-motion/<scene>/<subject>/<shard>.tar
└── EgoProMax/<scene>/<subject>/<shard>.tar

Each sample preserves a common key across media and annotations:

<sample_key>.mp4
<sample_key>.json
<sample_key>.jpg       # optional representative frame

Metadata contract

Field	Type	Description
sample_id	string	Globally unique anonymous sample identifier
sku	string	One of the six release SKUs
scene	string	Scene-family identifier
subject_id	string	Anonymous participant identifier
duration_sec	float	Clip duration in seconds
fps	float	Video frame rate
width, height	integer	Video resolution
views	list[string]	Available camera viewpoints
annotations	list[string]	Available supervision layers
sha256	string	File-integrity checksum

Access

The staging repository currently contains the Dataset Card and display assets. The following command tests repository access without referring to data files that have not yet been published:

hf download xuejf/EgoSuite-Open1.0K \
  --repo-type dataset \
  --include "README*" "assets/*" \
  --local-dir ./EgoSuite-Open10K-preview

Data-shard, manifest, streaming, and checksum examples will be activated after the audited production files are uploaded. The Dataset Viewer is intentionally disabled during the page-preview stage so display assets are not misidentified as dataset rows.

Data Provenance and Responsible Use

The release is derived from Lightwheel's real-environment egocentric human-data operations. The production version must document capture scope, hardware, annotation pipeline, participant authorization, anonymization, quality control, and exclusions at the same version as its audited manifest.

Users must comply with the final access terms, participant-consent scope, privacy requirements, and downstream-use restrictions. Users must not attempt to identify participants, reconstruct sensitive locations, or use the data for surveillance, profiling, or harmful applications.

License and Release Status

Item	Staging status
Dataset Card and visual assets	Available for review
Dataset files	Not published in this test repository
Audited release manifest	Pending final data-team lock
License and access terms	Pending Lightwheel legal approval
Official support email	Pending company confirmation

This staging repository does not grant a license to training data. The approved license and, if required, gated-access terms must be published before the dataset files become available.

Citation

@dataset{lightwheel_egosuite_open10k_2026,
  author    = {{Lightwheel}},
  title     = {EgoSuite-Open10K},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/xuejf/EgoSuite-Open1.0K}
}

The final citation will use the approved author list and official release URL.

Contact

For product information and collaboration, visit the Lightwheel EgoSuite product page. The release-specific support email will be added after company approval.

---

EgoSuite — scalable, structured human demonstrations for embodied intelligence.