Ego-DailyManipulation-EgoKit

English | 中文

English

Multimodal human demonstration data collected with Ego Kit and published by 北京行知启航科技有限公司 (XzqhTechLtd). The dataset covers tabletop sorting, garment folding, box packing, and box opening and closing. It includes dual RGB and four fisheye video streams, tactile glove measurements, and hand joint poses for gesture and motion trajectory analysis.

Capture equipment

Item	Specification
Device model	Ego Kit
Physical cameras	6: 2 RGB cameras + 4 fisheye cameras
Video frame rate	30 FPS
RGB resolution	1600 × 1200
Fisheye video resolution	640 × 480
Six-camera field of view	HFOV 118°, VFOV 94.8°
Four-camera fisheye field of view	DFOV 150°, HFOV 128°, VFOV 80°
IMU sampling rate	500 Hz (device specification)
Hand sensing	Tactile gloves, hand joints, gesture poses, and trajectories

The field of view and IMU sampling rate are nominal device specifications supplied by the publisher, rather than measurements derived from this dataset. The six-camera field-of-view entry describes the six-camera configuration and should not be interpreted as the field of view of every camera. Published frame-level observations are aligned at 30 FPS and include glove IMU acceleration and angular velocity fields; they do not constitute a separate raw 500 Hz IMU stream.

Dataset contents

The dataset uses LeRobot v3.0 and contains 5 episodes and 9,073 frames, totaling approximately 5 minutes.

Modality	Contents
RGB video	observation.images.rgb1, observation.images.rgb2
Fisheye video	observation.images.fe1 through observation.images.fe4
Hand joints and poses	3D coordinates for 21 joints per hand, 2D coordinates in both RGB views, and tracking validity flags
Tactile gloves	Bend measurements, pressure summaries, pressure points, IMU acceleration, and angular velocity
Trajectories	Hand joint poses and timestamps across consecutive frames for motion trajectory reconstruction and analysis
Task annotations	Task indices, semantic action labels, and segment indices
Observation plots	observation.state: 24 dimensions of glove bend and pressure summary values

observation.state concatenates flattened observation.glove_bend [2,5] and observation.glove_pressure_summary [2,7], preserving the source values and ordering. See state_schema.json for dimension definitions. This dataset records human hand observations and does not include robot control commands in an action field; semantic action labels describe the demonstrated activities.

Trajectories here refer to hand joint coordinate sequences. A separate device localization or global motion trajectory is not provided.

Recorded tasks

• Tabletop region sorting
• Garment folding
• Mailer box packing
• Box opening and closing
• Camera test

Visualization

Select an episode in Episodes Preview to view 6 camera streams, or use the official LeRobot Visualizer to explore the videos and glove observation plots.

Stereo hand skeleton replay videos are also provided as derived visualization attachments and are not counted among the 6 physical cameras. File locations for each episode are listed in derived_visualizations.json.

Timing and validity

Videos and frame-level observations are aligned at 30 FPS. Videos use H.264 / yuv420p MP4 encoding. Source timestamps, frame indices, glove time offsets, and keypoint validity flags are retained for temporal alignment and tracking quality analysis. This frame alignment preserves the supplied indexing and does not establish independently verified hardware synchronization. Hand keypoints include missing values (NaN); use the validity flags when analyzing joint trajectories. Glove values retain the source scale; physical units and left/right hand mapping have not been independently established. Refer to state_schema.json for the channel ordering.

---

中文

Ego 系列 · 日常操作 · EgoKit 数据集

北京行知启航科技有限公司（XzqhTechLtd）使用 Ego Kit 采集的多模态人类操作数据，涵盖桌面整理、衣物折叠、装盒及盒子开合等任务。数据包含双 RGB、四鱼眼视频、触觉手套信息和手部关节位姿，支持手势与运动轨迹分析。

采集设备

项目	参数
设备型号	Ego Kit
摄像机总数	6 个：2 个 RGB 相机 + 4 个鱼眼相机
视频帧率	30 FPS
RGB 分辨率	1600 × 1200
鱼眼视频分辨率	640 × 480
六目视场参数	HFOV 118°，VFOV 94.8°
四目鱼眼视场参数	DFOV 150°，HFOV 128°，VFOV 80°
IMU 采样频率	500 Hz（设备参数）
手部信息	触觉手套、手部关节、手势位姿与轨迹

视场及 IMU 采样频率为发布方提供的设备标称参数，并非从本数据集测量得出。“六目视场参数”对应六目配置，不应理解为每个相机均具有该视场。当前发布的帧级观测按 30 FPS 对齐，包含手套 IMU 加速度、角速度字段，不代表独立的 500 Hz 原始 IMU 数据流。

数据内容

数据采用 LeRobot v3.0 格式，包含 5 个 episode、9,073 帧，共约 5 分钟。

模态	内容
RGB 视频	observation.images.rgb1、observation.images.rgb2
鱼眼视频	observation.images.fe1 至 observation.images.fe4
手部关节与位姿	双手 21 关节三维坐标、双 RGB 视角二维坐标及跟踪有效性标记
触觉手套	弯曲、压力汇总、压力点、IMU 加速度与角速度
轨迹	连续帧的手部关节位姿及时间戳，可用于重建和分析运动轨迹
任务标注	任务索引、动作语义标签与片段索引
观测曲线	observation.state：24 维手套弯曲与压力汇总值

observation.state 依次展开 observation.glove_bend [2,5] 和 observation.glove_pressure_summary [2,7]，保留源数值及顺序。维度定义见 state_schema.json。本数据集记录人手观测，不包含机器人控制命令 action；动作语义标签用于描述操作内容。

此处“轨迹”指手部关节坐标序列；数据集未提供独立的设备定位或全局运动轨迹。

采集任务

• 桌面分区整理
• 衣物折叠
• 包装纸盒
• 盒子开合
• 摄像头测试

可视化

在 Episodes Preview 中选择 episode，查看 6 路相机视频；也可使用官方 LeRobot Visualizer查看视频及手套观测曲线。

另提供双目手部骨架回放视频，作为派生可视化附件，不计入 6 个实体摄像机。各 episode 的文件位置见 derived_visualizations.json。

时间与有效性

视频及帧级观测以 30 FPS 对齐，视频为 H.264 / yuv420p MP4。数据保留源时间戳、帧索引、手套时间差和关键点有效性标记，便于进行时间对齐与跟踪质量分析。帧级对齐沿用源索引关系，不代表已独立验证设备硬件同步。手部关键点包含缺失值（NaN），使用关节轨迹时应结合有效性标记。手套数值保留源量纲，物理单位和左右手映射尚未独立确认；通道顺序见 state_schema.json。