100,000+ HJ long videos in first-person perspective with head IMU at 200Hz, gaze point and hand keypoint annotations for VLA model pre-training
| Dataset Name | EgoSuite Egocentric Human Activity Dataset |
| Total Data Volume | 100,000+ Hours |
| Data Type | RGB-D video, IMU at 200 Hz, eye gaze, 21 hand keypoints, natural language commands |
| Video Specifications | 1920×1080 @ 30 FPS |
| Delivery Format | LeRobot 3.0 / RLDS / HDF5 / JSON Metadata |
| License Type | Licensed Data or Custom Projects |
Each piece of data contains complete structured fields, covering all requirements from basic metadata to professional analysis.
| Field Category | Typical Fields |
|---|---|
| Ego-centric Video | Video file, frame, resolution, recording time, camera calibration parameters, exposure and white balance, FOV 160° |
| Head Posture and IMU | 200 Hz velocity, gyroscope, magnetic field, head orientation, timestamp alignment |
| # Hand Keypoints # | 2D/3D gaze point, right-hand 21 keypoints, gesture recognition, operating object bounding box, exposure state |
| Action and Scene Labels | Action segment start time, scene tag, task goal, natural language command, failure/success marker, operating object state transition |
| Metadata and Compliance | Device model, collector role, privacy de-identification marker, authorization scope, data version |
The following is a de-identified structured JSON sample illustrating the typical data format.
Other Organiser
"EGO-2026-0089123",
"scene": "kichen morning",
"Duration sec": 1847,
"video", "video 0089123.mp4, "fps": 30, "resolution": [200, 1080],
"imu": "Path": "imu 0089123.csv", "sample rate hz": 200, "axes": ["acc x", "acc z", "gyro x", "gyro y", "gyro y", "gyro z"]
"gaze": {"timetestamps": [0.033, 0.066], "points 2d": [512, 384], [520, 390]},
"hand keypoints": [x,y,v], "right": [x,y,v],
"actions":
"Start": 12.5, "end": 18.2, "label": "open fridge", "object": "fridge", "success": true}
I don't know.
"nl instruation": "Take the milk out of the fridge and pour it into the cup."
♪ I'm sorry ♪Large-scale egocentric video paired with natural language commands, training vision-language-action foundation models.
First-person operating segments provide cross-scene action demonstrations for robotic arms/humanoid robots.
Continuous long-form video and multimodal signals build predictive world models of human daily activity.
Supports baseline tasks such as future action prediction, procedure learning, and object interaction testing.
Intent recognition and hand interaction models for head-worn devices, trained on observation and gesture data.
Home-scene video supports elderly living monitoring and smart assistant decision-making.
Changsha Langhui Information Technology Co., Ltd. The DataAssetsAPI platform is dedicated to providing high-quality, compliant data assets for AI companies and research institutions. It supports flexible combination by dimension and by quantity scale.
Data Sample Preview · Customized Data Solutions · Professional Technical Support