xuejf/EgoSuite-Open2.0K

EgoSuite-Open10K

Real-world human activity, structured for embodied intelligence.

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

Overview · Data SKUs · Scenes · Data contract · Access · 中文 · EgoSuite

Staging repository. xuejf/EgoSuite-Open2.0K is used to review the Dataset Card and visual system. The product shown on this page remains EgoSuite-Open10K, with a planned 10,000-hour portfolio. Dataset files, the audited manifest, access terms, and the official release URL are not published in this test repository.

Overview

EgoSuite-Open10K is Lightwheel's open-data program for embodied AI, robot learning, and world-model research. It organizes real-world egocentric human activity as a capability ladder: teams can select the camera viewpoints and supervision depth required by a task while retaining one shared data contract.

10,000 h	6	7	2	3
Planned portfolio	Data SKUs	Environment families	Camera viewpoints	Supervision layers

Designed around real-world learning

Design principle	Coverage	Research value
Human-centered activity	Natural routines and operational workflows	Captures long-horizon, compositional behavior beyond scripted lab demonstrations
Egocentric capture	Head RGB across every SKU	Preserves hand-object interaction, attention context, and temporal continuity
Multi-view capture	Synchronized wrist RGB in the Pro series	Adds detail for fine manipulation and reduces hand-object occlusion
Structured supervision	3D hand pose, 3D body pose, and V7 semantics	Supports perception, action understanding, imitation learning, and multimodal reasoning

Six Data SKUs

The Standard series scales head-view capture. The Pro series adds a synchronized wrist viewpoint. Motion and fully supervised tiers progressively add full-body pose and semantic labels.

Series	SKU	Head RGB	Wrist RGB	3D hand pose	3D body pose	V7 semantics	Planned hours
Standard	EgoStandard	✓	—	✓	—	—	8,400
Standard	EgoStand-motion	✓	—	✓	✓	—	500
Standard	EgoFull	✓	—	✓	✓	✓	100
Pro	EgoPro	✓	✓	✓	—	—	750
Pro	EgoPro-motion	✓	✓	✓	✓	—	200
Pro	EgoProMax	✓	✓	✓	✓	✓	50
						Total	10,000

SKU selection guide

• Choose EgoStandard when scale and head-view hand-object interaction matter most.
• Choose a motion tier for body dynamics, action trajectories, or whole-body understanding.
• Choose the Pro series when fine manipulation benefits from a synchronized wrist view.
• Choose EgoFull or EgoProMax for the deepest pose and semantic supervision.

The OPEN markers in the visual identify the current candidate fully supervised tiers for this release review. Final availability must follow the audited release manifest.

Seven Environment Families

Environment	Representative activities
Home	Cooking, cleaning, daily routines, and object organization
Hospitality	Food preparation, service workflows, and table operations
Retail	Picking, restocking, checkout, and customer-service workflows
Sports	Training, equipment handling, and coordinated movement
Logistics	Sorting, packing, transport, and warehouse operations
Office	Device use, document handling, desktop work, and collaboration
Industry	Tool use, assembly, inspection, and industrial operations

Dataset at a Glance

The taxonomy is hierarchical rather than flat:

Environment family
└── Scene type
    └── Human-centered task
        └── Synchronized media + annotations

This structure supports broad environment filtering, fine-grained task selection, and cross-scene evaluation without changing sample-level keys.

Data Contract

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

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

One anonymous sample key links every available view and supervision layer:

<sample_key>.head.mp4
<sample_key>.wrist.mp4          # Pro series only
<sample_key>.hand_pose.parquet
<sample_key>.body_pose.parquet  # motion / full tiers
<sample_key>.semantics.jsonl    # Full / ProMax
<sample_key>.metadata.json

Core metadata

Field	Type	Description
sample_id	string	Globally unique anonymous sample identifier
sku	string	One of the six EgoSuite SKUs
environment	string	Top-level environment family
scene	string	Fine-grained scene identifier
task	string	Human-centered task label
subject_id	string	Anonymous participant identifier
duration_sec	float	Clip duration in seconds
views	list[string]	Available camera viewpoints
annotations	list[string]	Available supervision layers
sha256	string	File-integrity checksum

The production manifest will define the authoritative schema, splits, counts, durations, and checksums for each release version.

Access and Release Status

The staging repository currently contains only the Dataset Card and display assets. Use the following command to verify the new repository after upload:

hf download xuejf/EgoSuite-Open2.0K \
  --repo-type dataset \
  --include "README*" "assets/*" \
  --local-dir ./EgoSuite-Open10K-card-preview

Item	Staging status
Dataset Card and four visual assets	Ready for review
Dataset shards	Not published in this test repository
Audited manifest and checksums	Pending data-team lock
License and access terms	Pending Lightwheel approval
Official support email	Pending company confirmation

The Dataset Viewer is intentionally disabled during page review so that display images are not interpreted as dataset rows. Streaming and selective-download examples should be activated only after the real shard paths and schema are locked.

Data Provenance and Responsible Use

The planned release is derived from Lightwheel's real-environment egocentric human-data operations. The production Dataset Card must document capture hardware, participant authorization, anonymization, annotation provenance, quality control, excluded content, and known limitations at the same version as the audited manifest.

Users must follow the final license, consent scope, privacy requirements, and downstream-use restrictions. Prohibited uses should include participant re-identification, sensitive-location reconstruction, surveillance, profiling, and harmful applications.

License

This staging repository does not grant a license to training data. The approved dataset license and, if required, gated-access conditions must be added before production files are made available.

Citation

@dataset{lightwheel_egosuite_open10k_2026,
  author    = {{Lightwheel}},
  title     = {EgoSuite-Open10K},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/xuejf/EgoSuite-Open2.0K}
}

The official release must replace the staging URL and use the company-approved author list.

Contact

For product information and collaboration, visit the Lightwheel EgoSuite product page. The dataset support email will be added after company confirmation.

---

EgoSuite — scalable, structured human demonstrations for embodied intelligence.