Egocentric-100K — Manual Labor Egocentric Video Dataset
Egocentric-100K is the largest dataset of manual labor egocentric video, containing 100,405 hours of first-person recordings from head-mounted fisheye cameras. Released in February 2026 under Apache 2.0 for commercial use, it features state-of-the-art hand visibility and active manipulation density specifically designed for training robotic manipulation policies. The head-mounted fisheye camera configuration provides wider field of view than standard wrist cameras, capturing both hands and the surrounding environment simultaneously. The dataset covers industrial, service, and domestic manual labor tasks across diverse real-world settings. At 100K hours it is the largest single-source egocentric video dataset available for robot learning.
| Year | 2026 |
|---|---|
| Total hours | 100,405 |
| Embodiments | human (head-mounted fisheye camera) |
| Modalities | rgb, proprioception |
| Task categories | manipulation, cleaning, cooking, pick-and-place, inspection, human-robot-interaction |
| Data format | mp4, json |
| License | Apache 2.0 |
| Access | gated — commercial use permitted |
| Maintainer | Build AI |
| Origin country | US |
What is it?
Egocentric-100K is the largest dataset of manual labor egocentric video, containing 100,405 hours of first-person recordings from head-mounted fisheye cameras. Released in February 2026 under Apache 2.0, it was specifically designed for training robotic manipulation policies with state-of-the-art hand visibility and active manipulation density. The head-mounted fisheye configuration provides wider field of view than standard wrist cameras, capturing both hands and surrounding environment simultaneously across industrial, service, and domestic manual labor tasks.
Who is it for?
Researchers training robotic manipulation policies that require rich hand-object interaction data at scale. The manual labor focus covers exactly the task categories most relevant to near-term robot deployment — warehouse operations, food service, cleaning, assembly — with real-world diversity that lab datasets cannot match.
Key specifications
- Duration: 100,405 hours of egocentric video
- Camera: Head-mounted fisheye
- Focus: Manual labor tasks — industrial, service, domestic
- Modalities: RGB (fisheye), proprioception
- License: Apache 2.0 — commercial use permitted
- Access: Open — Hugging Face
How it compares
Egocentric-100K is the largest single-source egocentric video dataset — 10x larger than Ego4D (3,025 hours) and 80x larger than EgoVerse (1,213 hours). Its manual labor focus and commercial license make it the strongest pretraining source for service robot policies. The fisheye camera configuration differs from standard pinhole cameras used in most datasets, providing wider coverage but requiring calibration for policy transfer.
Limitations and access notes
Manual labor focus means less coverage of fine dexterous manipulation compared to task-specific datasets like DexWild. Fisheye distortion requires correction or fisheye-aware model architectures. Apache 2.0 permits unrestricted commercial use.
Linked professions
- Hotel Housekeeper
- Warehouse Picker Packer
- Fast Food Worker
- Commercial Floor Cleaner
- Laundry Worker Commercial
- Assembly Line Worker Repetitive
Frequently asked questions
What makes Egocentric-100K unique?
Egocentric-100K is the largest egocentric video dataset focused on manual labor at 100,405 hours — 10x larger than Ego4D. It uses head-mounted fisheye cameras providing wider field of view than wrist cameras, capturing both hands and environment simultaneously. The manual labor focus covers industrial, service, and domestic tasks directly relevant to robot deployment.
Can Egocentric-100K be used commercially?
Yes. Licensed under Apache 2.0, permitting unrestricted commercial use, modification, and redistribution.
What is 'active manipulation density'?
Active manipulation density refers to the proportion of recorded time where hands are actively interacting with objects. Egocentric-100K was designed to maximize this ratio — minimizing passive walking or observation time and maximizing frames containing hand-object interaction, which is the signal most useful for training manipulation policies.
How do I access Egocentric-100K?
Available on Hugging Face at huggingface.co/datasets/EgocentricVision/Egocentric-100K. No registration required.
Why fisheye cameras instead of standard cameras?
Fisheye cameras provide 180+ degree field of view versus ~60-90 degrees for standard cameras. For egocentric manipulation, this captures both hands, nearby objects, and the wider environment simultaneously — providing richer spatial context than narrow field of view cameras that often miss one hand or the surrounding scene.