光轮智能/EgoDemo
数据集
数据集查看器
文件和版本
Pull Requests
讨论

简体中文 | English

光轮智能

EgoDemo

覆盖全部四个 SKU 的 50 小时样本数据集,建议从这里开始。

所有参与者均已提供知情同意,涵盖数据采集、标注、受控数据集发布,以及仓库许可证所允许的用途。

EgoSuite-Open100K 概览

Collection: EgoSuite-Open100K
SKUSub-SKUFormatPlanned Duration
EgoStandardEgoStandHand Pose80,000 h
EgoStand-bodyHand Pose + Full Body Pose10,000 h
EgoProEgoProStandardWrist + Hand Pose8,000 h
EgoProStandard-bodyWrist + Hand Pose + Full Body Pose2,000 h
EgoDemoFour SKUs + Two Raw VideosRaw Video50 h

EgoSuite-Open100K 是一个分阶段发布的 100,000 小时开放数据集,兼顾标注质量、规模和多样性。数据覆盖 15,000+ 项任务和 15,000+ 个不同场景,提供手部与全身姿态;部分数据包含腕部相机视角,并额外提供子任务级语义标注。

EgoDemo 数据集摘要

EgoDemo 是从 EgoStandard 和 EgoPro 中抽取的 50 小时样本,覆盖全部四个已标注子集和两个原始视频变体。上文中的 15,000+ 项任务及场景数据描述的是完整合集,而非 EgoDemo 本身。同一批样本同时提供 LeRobot v3 和 MCAP 格式。

包含的变体

Source SKUVariantCore contentSummary
EgoStandardEgoStandHead-view video + hand poseA high-volume head-view sample for hand-centric activity.
EgoStandardEgoStand-bodyHead-view video + hand pose + full-body poseAdds whole-body motion context to head-view hand interaction.
EgoProEgoProStandardHead-view and wrist-view video + hand poseAdds a close-range wrist view for contact and grasp.
EgoProEgoProStandard-bodyHead-view and wrist-view video + hand pose + full-body poseCombines close-range hand detail with whole-body motion.
EgoDemoEgoRawRaw head-view videoA pose-free raw head-view sample for lightweight inspection.
EgoDemoEgoProRawRaw head-view and wrist-view videoA pose-free synchronized head-and-wrist raw sample.

EgoRaw 和 EgoProRaw 是 EgoDemo 独有的原始视频变体。它们是独立交付变体,不属于 EgoStandard 或 EgoPro 的标准已标注 Sub-SKU。

标注模态

ModalityEgoStandEgoStand-bodyEgoProStandardEgoProStandard-bodyEgoRawEgoProRaw
Head-view videoYesYesYesYesYesYes
Wrist-view videoNoNoYesYesNoYes
Left/right hand poseYesYesYesYesNoNo
Full-body poseNoYesNoYesNoNo
Standard SKU includes semanticsNoNoNoNoNot applicableNot applicable
Complimentary semantic annotations in this releaseYesYesYesYesNoNo

已标注变体中的姿态值使用 float32 类型和世界坐标。手部位置数组的 shape 为 (21, 3),手部旋转为 (21, 4),全身位置为 (22, 3),全身旋转为 (22, 4)。四元数值遵循 (qw, qx, qy, qz) 顺序。

四个已标注变体使用两类标准标注:手部姿态,以及 body 变体中的全身姿态。两个原始视频变体不包含姿态标注。在本次开源发布中,光轮智能为四个已标注变体免费附赠时序事件级语义标注;该附加标注不适用于原始视频变体。

可用格式与 Episode 布局

FormatRecommended useDirectoryEpisode contents and documentation
LeRobot v3.0Annotated samples and LeRobot-compatible workflowsEgoStand/lerobot/, EgoStand-body/lerobot/, EgoProStandard/lerobot/, EgoProStandard-body/lerobot/Per-episode Parquet, video, metadata, and release annotations; see the Sub-SKU README files
MCAPAnnotated synchronized multimodal samplesEgoStand/mcap/, EgoStand-body/mcap/, EgoProStandard/mcap/, EgoProStandard-body/mcap/Per-episode MCAP package and verified sidecar files; see the Sub-SKU README files
MP4Two EgoDemo-only Raw Video variantsEgoRaw/mp4/, EgoProRaw/mp4/MP4 video files; variant manifests are stored beside each Raw Video directory

数据集结构

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/

每个变体均位于各自的 Sub-SKU 目录下。若某个格式目录不存在,则表示该变体不提供相应格式的交付内容。

所有已发布数据均使用上述 Sub-SKU 优先路径。两个原始视频变体分别使用 EgoRaw 和 EgoProRaw 目录名称。

Manifest 数据结构

已标注 EgoDemo 格式的 manifest 保留来源仓库、来源 Sub-SKU、来源 Episode UUID、不可变来源版本、格式路径、模态标志及校验和。同一个 source_episode_uuid 用于关联 LeRobot 和 MCAP 两种表示,不会被计为不同样本。

原始视频 manifest 记录以 Lightwheel/EgoDemo 作为来源,不包含来源 Sub-SKU、姿态模态或标注版本。

FieldDescription
demo_episode_idStable EgoDemo episode identifier
source_repoSource repository
source_skuSource SKU
source_sub_skuSource Sub-SKU for annotated variants; null for Raw Video
source_episode_uuidOriginal episode UUID used for cross-format matching
source_revisionImmutable source revision
variantAnnotated Sub-SKU name or Raw Video directory label
task_nameNormalized task name
scene_idAudited scene identifier
scene_familyTop-level scene family
duration_sAudited usable duration in seconds
episode_pathRepository-relative path for this format
formatlerobot-v3.0, mcap, or mp4-json
has_wrist_videoWhether wrist-view video is present
has_hand_poseWhether hand pose is present
has_body_poseWhether full-body pose is present
annotation_versionInternal schema identifier for annotated release samples; null for Raw Video
annotation_distributionComplimentary add-on marker when applicable
sha256Integrity checksum

发布版本

Release itemValue
Dataset release versionv1.0.0
LeRobot formatv3.0
MCAP containerMCAP0
MCAP schema encodingEmbedded Protobuf schemas
MCAP writer librarymcap-python 1.4.0
Raw Video formatMP4 + JSON

按需下载

git clone https://ai.atomgit.com/Lightwheel/EgoDemo.git
cd EgoDemo
git lfs pull --include="*/lerobot/**"

下载已标注 MCAP 数据时使用 */mcap/**;下载原始视频时使用 EgoRaw/mp4/** 或 EgoProRaw/mp4/**。