README.md

metadata

pretty_name: EgoSuite-Open100K - EgoPro
language:
  - en
license: other
license_name: commercial-training-no-resale-v1.0
size_categories:
  - 10K<n<100K
task_categories:
  - video-classification
tags:
  - video
  - egocentric-video
  - embodied-ai
  - human-demonstration
  - human-pose
  - hand-pose
  - body-pose
  - wrist-camera
  - multimodal
  - lerobot
  - mcap
  - robotics
gated: true
viewer: false
extra_gated_prompt: >-
  To help Lightwheel understand how this dataset is used, please provide the
  information below. Access covers both the LeRobot and MCAP directories in this
  repository and remains subject to the dataset license and usage terms.
extra_gated_fields:
  Full name: text
  Work email: text
  Organization / affiliation: text
  Role: text
  Country or region: text
  Intended use: text
  Research or product area: text
  Preferred data format (LeRobot, MCAP, or Both): text
  Commercial or non-commercial use: text
  I agree to the dataset license and usage terms: checkbox

EgoPro

10,000 hours of head-and-wrist egocentric human activity data

All participants provided informed consent covering data collection, annotation, controlled dataset release, and the uses permitted by the repository license.

Human Data Overview

Collection
🚀 EgoSuite-Open100K
Data Repositories
EgoStandard	EgoPro ⭐ CURRENT DATASET	EgoDemo
🌐 EgoSuite Website

Dataset Summary

EgoPro is a synchronized head-and-wrist egocentric human activity dataset with 3D hand pose and an optional full-body pose modality. It is distributed in both LeRobot and MCAP formats within this repository.

Collection Scale

The Lightwheel open-source egocentric collection spans 15,000+ tasks and 15,000+ scenes. These are collection-level coverage figures shared across the release and should not be added across repositories or delivery formats.

EgoPro contains 10,000 hours of data:

Sub-SKU	Duration	Standard SKU definition
EgoProStandard	8,000 hours	Head-view and wrist-view video + left/right hand pose
EgoProStandard-body	2,000 hours	Head-view and wrist-view video + left/right hand pose + full-body pose

LeRobot and MCAP are alternative representations of these same 10,000 hours and must not be added together.

Sub-SKUs

The duration breakdown and standard definitions are listed in the Collection Scale table above.

Complimentary Semantic Annotations

The standard EgoPro Sub-SKUs are defined by their video and pose modalities and do not include semantic annotations.

For this open-source release, Lightwheel provides temporal semantic annotations as a complimentary add-on. Their presence in the released files must not be interpreted as part of the standard EgoPro SKU specification or as a commitment that semantic annotations are included in future standard deliveries.

The underlying release files are unchanged: the semantic annotations remain available in the format-specific representation documented below.

Available Formats

Format	Recommended use	Location	Documentation
LeRobot v3.0-compatible	Training and LeRobot-compatible robotics pipelines	lerobot/	lerobot/README.md
MCAP	Synchronized multimodal streams and MCAP-compatible workflows	mcap/	mcap/README.md

Dataset Structure

EgoPro/
├── README.md
├── lerobot/
│   ├── README.md
│   ├── data/
│   │   ├── EgoProStandard/
│   │   └── EgoProStandard-body/
│   └── manifests/
└── mcap/
    ├── README.md
    ├── data/
    │   ├── EgoProStandard/
    │   └── EgoProStandard-body/
    └── manifests/

Episode Layout

Episode packaging differs by format:

Format	Episode layout
LeRobot	Per-episode Parquet, synchronized head/wrist video, LeRobot metadata, and release annotation files; see lerobot/README.md
MCAP	Per-episode MCAP package and any verified sidecar files; see mcap/README.md

Modalities

Modality	EgoProStandard standard SKU	EgoProStandard-body standard SKU	Included in this open-source release
Head-view egocentric video	Yes	Yes	Yes
Wrist-view egocentric video	Yes	Yes	Yes
Left/right 3D hand pose	Yes	Yes	Yes
Full-body 3D pose	No	Yes	Body variant only
Semantic annotations	No	No	Yes, complimentary add-on
Frame-quality metadata	Delivery-dependent	Delivery-dependent	Included when available

Pose values use float32 and world coordinates. Hand position arrays have shape (21, 3), hand rotations have shape (21, 4), full-body positions have shape (22, 3), and full-body rotations have shape (22, 4). Quaternion values follow (qw, qx, qy, qz) ordering.

When per-frame wrist-camera 6DoF pose is provided, clients should use the feature or topic names declared by the selected format instead of inferring them from the video-channel name.

Manifest Schema

LeRobot and MCAP are alternative representations of the release. The format-specific manifests use the same stable episode_uuid and sub_sku values to identify corresponding episodes. Users must not add LeRobot and MCAP rows together when calculating dataset scale.

Common manifest fields are:

Field	Description
episode_uuid	Stable episode identifier shared across formats
human_case_id	Stable Human Case identifier
sku	Always EgoPro
sub_sku	EgoProStandard or EgoProStandard-body
task_name	Normalized task name
scene_id	Audited scene identifier
scene_family	Top-level scene family
duration_s	Audited usable duration in seconds
episode_path	Repository-relative path for this format
format	lerobot-v3.0 or mcap
schema_version	Format-specific schema/profile version
has_wrist_video	Whether wrist-view video is present
has_hand_pose	Whether hand pose is present
has_body_pose	Whether full-body pose is present
annotation_version	Internal schema identifier when the complimentary semantic annotation is present
annotation_distribution	complimentary-open-source-add-on when applicable
source_revision	Immutable source revision
sha256	Integrity checksum

Release Versions

Release item	Value
Dataset release version	[VERIFY]
LeRobot format	v3.0-compatible
MCAP version/profile	[VERIFY]
MCAP schema	[VERIFY]
Converter version	[VERIFY]
Semantic annotations	Included as a complimentary release add-on; internal version not exposed
Annotation distribution	Complimentary open-source release add-on

Selective Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="LightwheelAI/EgoPro",
    repo_type="dataset",
    allow_patterns=["lerobot/**"],
)

Download MCAP by replacing lerobot/** with mcap/**.