Dataset
EgoDemo is the developer-friendly entry point to EgoSuite-Open100K. Unlike EgoStandard and EgoPro, its data is stored directly in this dataset repository. The annotated samples are available in both LeRobot v3 and MCAP; the raw-video samples are MP4.
The 15,000+ task and scene figures associated with EgoSuite-Open100K describe the full collection, not this 50-hour sample.
Accept the access terms on this page, then authenticate with the Hugging Face CLI:
pip install -U huggingface_hub
hf auth login
Download the complete repository:
hf download LightwheelAI/EgoDemo \
--repo-type dataset \
--local-dir ./EgoDemo
Use --include to download only the data you need:
# One annotated LeRobot variant
hf download LightwheelAI/EgoDemo \
--repo-type dataset \
--include "EgoStand/lerobot/**" \
--local-dir ./EgoDemo
# Raw head-view video
hf download LightwheelAI/EgoDemo \
--repo-type dataset \
--include "EgoRaw/mp4/**" \
--local-dir ./EgoDemo
See the Hugging Face download documentation for caching, filtering, and dry-run options.
EgoDemo/
├── README.md
├── EgoStand/
│ ├── README.md
│ ├── lerobot/
│ ├── mcap/
│ └── manifests/
├── EgoStand-body/
│ ├── README.md
│ ├── lerobot/
│ ├── mcap/
│ └── manifests/
├── EgoProStandard/
│ ├── README.md
│ ├── lerobot/
│ ├── mcap/
│ └── manifests/
├── EgoProStandard-body/
│ ├── README.md
│ ├── lerobot/
│ ├── mcap/
│ └── manifests/
├── EgoRaw/
│ ├── README.md
│ ├── mp4/
│ └── manifests/
└── EgoProRaw/
├── README.md
├── mp4/
└── manifests/
An absent format directory means that format is not provided for the variant. Variant-specific details are documented in the README inside each directory.
LeRobot and MCAP are alternative representations of the same annotated episodes. The shared source_episode_uuid connects formats; do not add their durations or episode counts together.
Pose data uses float32 world coordinates:
| Value | Shape |
|---|---|
| Left/right hand positions | (21, 3) |
| Left/right hand rotations | (21, 4) |
| Full-body positions | (22, 3) |
| Full-body rotations | (22, 4) |
Quaternions use (qw, qx, qy, qz) ordering. Temporal event-level semantic annotations are included for the four annotated variants; the raw-video variants contain no pose or semantic annotations.
Manifest rows provide stable identifiers, source provenance, task and scene metadata, duration, repository-relative paths, modality flags, format information, and SHA-256 checksums. Raw-video manifest rows omit source Sub-SKU and pose fields.
| Item | Version |
|---|---|
| Dataset release | v1.0.0 |
| LeRobot | v3.0 |
| MCAP container | MCAP0 |
| MCAP schema encoding | Embedded Protobuf |
| Raw video | MP4 + JSON |
@misc{lightwheel2026egosuite,
author = {{Lightwheel}},
title = {{EgoSuite-Open100K: A Fully Annotated Open Egocentric Human Dataset}},
year = {2026},
month = aug,
howpublished = {\url{https://huggingface.co/collections/LightwheelAI/egosuite-open100k}},
note = {Dataset}
}
All participants provided informed consent covering collection, annotation, controlled release, and uses permitted by the dataset license. Recordings were de-identified before release. Use of the data is governed by the repository license and access terms.
For dataset questions, open a discussion in this repository. For other inquiries, contact contact@lightwheel.ai.