Human Egocentric Video Dataset for Embodied AI and Robotics
Stereo egocentric video with synchronized head and hand pose, captured on mixed-reality hardware and ready for your training pipeline today.
Explore the Dataset
Conference Room
Kitchenette Pantry
Open Office Desk
Reception Desk
Our egocentric dataset is available to license immediately, delivering production-grade stereo video, synchronized head and hand pose, and VLA-ready annotations from real human manipulation tasks captured across real-world environments.
- Rectified stereo video above 1080p with 21 to 26-joint hand tracking per hand
- Delivered in HDF5 and LeRobot v2 formats for immediate pipeline ingestion
- Commercial usage rights with PII removed at the point of capture
- Purposeful manipulation episodes, so every clip earns its place in your training set
Access Dataset
| Environment | Task | Frames | Duration | Download |
|---|---|---|---|---|
| Conference Room | Distribute notepads and pens around a conference table | 3,674 | 61.2 | |
| Conference Room | Erase a whiteboard and rewrite a 3-row agenda | 6,296 | 104.9 | |
| Conference Room | Set up a water carafe and six glasses on the conference table | 4,136 | 68.9 | |
| Conference Room | Set up six chairs around a conference table | 5563 | 93 | |
| Kitchenette Pantry | Heat a lunch container in a kitchenette microwave | 6,649 | 110 | |
| Kitchenette Pantry | Pour hot water from a dispenser for instant tea | 5350 | 107 | |
| Kitchenette Pantry | Wipe down a microwave interior with a damp paper towel | 4750 | 79 | |
| Open Office Desk | Refill a self-inking stamp with new ink drops | 3,734 | 62.2 | |
| Open Office Desk | Replace a dry whiteboard marker tip cap and store in tray | 4401 | 73 | |
| Reception Desk | Sign a visitor into a paper logbook and issue a badge | 3939 | 66 |
Explore our Egocentric Dataset Catalog
Robots learn manipulation best from real human demonstrations. Our ready-to-license, first-person data captures people doing real work, recorded with synchronized motion sensors, so your team can begin training right away instead of spending months building its own collection.
Industrial Workplace Video
Egocentric RGB footage of skilled work across nine real workplaces, including garment factories, furniture workshops, electronics and semiconductor labs, paper and textile mills, and household and retail settings.
iPhone Home Tasks
Everyday household tasks captured on iPhone, with depth, 3D hand tracking, IMU data, and camera calibration included in every clip.
Narrated Video with Transcripts
Workers describe what they are doing as they do it, and every clip comes with a transcript. This pairing of action and language is well suited to training vision-language-action models.
Calibrated Stereo
Calibrated global-shutter stereo video with a 300 Hz IMU, built for metric 3D reconstruction of hands and objects.
Stereo with Wrist Cameras
A head-level stereo view combined with a close-up camera on each wrist, capturing fine-grained detail of how hands interact with objects.
UMI Gripper Demonstrations
Handheld-gripper episodes recorded with the UMI interface, structured to map directly to robot actions.
Why Choose DDD Egocentric Dataset
Mixed-Reality Headset Hardware
Captured using Pico 4 Ultra, Apple Vision Pro, Meta Quest 3, and Project Aria, devices with on-board SLAM, inside-out tracking, and hardware-level hand and head pose pipelines built for highest quality and precision manipulation capture.
Stereo Video with Full Pose Telemetry
Every episode delivers rectified stereo video above 1080p at 60fps, per-frame 6DoF head pose, a 21 to 26-joint hand skeleton per hand in both camera and world frames, and IMU data. Frame-to-pose synchronization is within 10 milliseconds at p95 across every episode.
Pipeline-Ready Delivery
Every episode ships as rectified stereo MP4 files, a raw HDF5 pose file with the full per-frame schema, and a complete LeRobot v2 dataset with a 371-dimensional float32 observation state.
Annotated for the Reasoning Layer
Every episode carries a natural-language task description and temporal sub-task segments of one to five seconds, each naming the exact object and action involved.
