EgoSuite-Open1K

Real-world human activity, structured for embodied AI.

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

Dataset Overview · Product Matrix · Scene Coverage · Data Structure · Quick Start · Access

---

Dataset Overview

EgoSuite-Open1K is a large-scale egocentric dataset for embodied intelligence research. Instead of exposing one undifferentiated collection, the release is organized as six composable data SKUs with progressively richer viewpoint, pose, motion, and semantic supervision.

10,000 h	7	6	2 viewpoints
Open subset	Scene families	Research SKUs	Head + wrist

Design principle — Preserve real-world behavioral diversity while making each supervision layer explicit, comparable, and reproducible.

What makes the suite different

Dimension	Coverage	Research value
Real environments	Home, hospitality, retail, sports, logistics, office, industry	Transfer beyond controlled lab demonstrations
Egocentric video	Head-mounted video across all SKUs	Natural hand-object interaction and temporal context
Multi-view capture	Wrist view in EgoPro series	Fine-grained manipulation and occlusion recovery
Pose supervision	Hand pose; full-body pose in motion tiers	Perception, imitation learning, action understanding
Semantic supervision	V7 semantic layer in EgoFull and EgoProMax	Structured reasoning over actions, objects, and scenes

Visual Preview

Replace assets/demo.mp4 with the approved 60-second H.264 release video before launch. Keep the filename unchanged and the page will update automatically.

Product Matrix

EgoSuite separates the 10K-hour open release into a Standard series and a Pro series. Researchers can choose the minimum supervision layer required by their task.

Series	SKU	Viewpoint	Pose / motion	V7 semantics	Release hours
Standard	EgoStandard	Head	Hand pose	—	8,400
Standard	EgoStand-motion	Head	Hand + body pose	—	500
Standard	EgoFull	Head	Hand + body pose	✓	100
Pro	EgoPro	Head + wrist	Hand pose	—	750
Pro	EgoPro-motion	Head + wrist	Hand + body pose	—	200
Pro	EgoProMax	Head + wrist	Hand + body pose	✓	50
				Total	10,000

How to choose a SKU

• Start with EgoStandard for scalable video pretraining and egocentric perception.
• Use motion tiers when full-body motion or action dynamics matter.
• Use the Pro series when head + wrist views are needed for fine manipulation.
• Use EgoFull or EgoProMax when semantic supervision is required.

Scene Coverage

The release spans seven scenario families so models can learn transferable interaction patterns across domestic and professional contexts.

01	02	03	04	05	06	07
Home	Hospitality	Retail	Sports	Logistics	Office	Industry
Daily routines	Service flows	Picking & checkout	Training	Sorting & packing	Desktop work	Tools & assembly

Data Modalities

Modality	Standard	Standard Motion	Full	Pro	Pro Motion	ProMax
Head-mounted RGB video	✓	✓	✓	✓	✓	✓
Wrist-view RGB video	—	—	—	✓	✓	✓
Hand pose	✓	✓	✓	✓	✓	✓
Full-body pose	—	✓	✓	—	✓	✓
V7 semantic annotation	—	—	✓	—	—	✓

Data Structure

The recommended production release uses WebDataset shards for large-scale streaming and reproducible partial download.

xuejf/EgoSuite-Open1K/
├── README.md
├── README_zh.md
├── assets/
│   ├── hero-egosuite-open10k.png
│   ├── demo-poster.png
│   └── demo.mp4
├── manifests/
│   ├── dataset_manifest.parquet
│   ├── shard_manifest.csv
│   └── checksums.sha256
├── EgoStandard/
│   └── <scene>/<subject>/<shard>.tar
├── EgoStand-motion/
├── EgoFull/
├── EgoPro/
├── EgoPro-motion/
└── EgoProMax/

Each shard should preserve a common sample key across media and annotations:

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

Suggested metadata schema

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

Quick Start

Stream with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset(
    "xuejf/EgoSuite-Open1K",
    streaming=True,
)

sample = next(iter(dataset["train"]))
print(sample.keys())

Download a specific file

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xuejf/EgoSuite-Open1K",
    filename="manifests/shard_manifest.csv",
    repo_type="dataset",
)

The loading code must be tested again after the final shard layout and dataset configuration are committed.

Responsible Use

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

Access & License

The final access policy and license must be confirmed by the dataset owner before public launch. If the release uses gated access, users will be required to sign in and accept the repository terms before downloading data files.

Citation

@dataset{egosuite_open10k_2026,
  author    = {{EgoSuite Team}},
  title     = {EgoSuite-Open1K},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/xuejf/EgoSuite-Open1K}
}

Contact

For access, research collaboration, or dataset issues, contact the official dataset team.

---

EgoSuite-Open1K — real-world egocentric data for embodied intelligence.