Egocentric-1M — The Internet for Physical AI
Egocentric-1M is the largest egocentric video dataset ever released — approximately 1 million hours of first-person factory and manual-labor footage from head-mounted fisheye cameras, released by Build AI on April 8, 2026 under Apache 2.0. It is the third and largest release in Build AI's scaling series: Egocentric-10K (10,000 hours, November 2025), Egocentric-100K (100,000+ hours, December 2025), and Egocentric-1M (~1 million hours, April 2026). Founder Eddy Xu described it as 'the internet for physical AI.' At roughly 1 million hours it exceeds Ego4D (3,025 hours) by more than 300x. The dataset focuses on real production environments — assembly lines, sorting, packaging, and machining — captured on custom head-mounted glasses with state-of-the-art hand visibility and active manipulation density. Access requires accepting conditions on Hugging Face; the Apache 2.0 license permits commercial use.
| Year | 2026 |
|---|---|
| Total hours | 1,000,000 |
| Embodiments | human (head-mounted fisheye camera) |
| Modalities | rgb, proprioception |
| Task categories | manipulation, pick-and-place, assembly, inspection, long-horizon |
| Data format | mp4, json, webdataset |
| License | Apache 2.0 |
| Access | gated — commercial use permitted |
| Maintainer | Build AI |
| Origin country | US |
What is it?
Egocentric-1M is the largest egocentric video dataset ever released — approximately 1 million hours of first-person recordings from head-mounted fisheye cameras, published by Build AI on April 8, 2026. It is the culmination of a rapid scaling series: Egocentric-10K (10,000 hours, 2,153 factory workers, 1.08 billion frames, November 2025), Egocentric-100K (100,000+ hours, December 2025), and Egocentric-1M (~1 million hours, April 2026). Build AI founder Eddy Xu described the dataset as 'the internet for physical AI' — a foundation-scale pretraining corpus for embodied intelligence.
Who is it for?
Teams pretraining robot foundation models, world models, and manipulation policies at scale. At 1M hours, Egocentric-1M is the first egocentric dataset large enough to support foundation-model pretraining from a single source. The factory and manual-labor domain focus makes it directly relevant to industrial manipulation and human-to-robot skill transfer.
Key specifications
- Duration: ~1,000,000 hours of egocentric video
- Camera: Head-mounted fisheye (OpenCV Kannala-Brandt fisheye model)
- Domain: Real factory and manual-labor environments — assembly, sorting, packaging, machining
- Format: 1080p 30fps, WebDataset shards with paired JSON metadata
- License: Apache 2.0 — commercial use permitted
- Access: Gated — requires accepting conditions on Hugging Face
- Released: April 8, 2026
How it compares
Egocentric-1M is more than 300x larger than Ego4D (3,025 hours) and roughly 10x larger than its predecessor Egocentric-100K. Xperience-10M (10,000 hours) offers richer per-session modalities such as motion capture and depth but at a fraction of the scale. The tradeoff is modality depth versus raw scale: Egocentric-1M is the scale play; Xperience-10M is the modality play.
Limitations and access notes
RGB fisheye video only — no depth, motion capture, or tactile modalities. Human data requires retargeting to robot kinematics. Fisheye distortion requires correction or fisheye-aware model architectures. The domain is factory and manual labor rather than domestic settings. Access requires Hugging Face login and condition acceptance; Apache 2.0 permits commercial use.
Linked professions
Frequently asked questions
How large is Egocentric-1M?
Approximately 1 million hours of egocentric video — the largest egocentric dataset ever released, exceeding Ego4D (3,025 hours) by more than 300x. Build AI released it on April 8, 2026.
How does Egocentric-1M relate to Egocentric-100K and Egocentric-10K?
All three are from Build AI using the same head-mounted fisheye camera configuration and factory manual-labor focus. Egocentric-10K (10,000 hours) launched in November 2025, Egocentric-100K (100,000+ hours) in December 2025, and Egocentric-1M (~1 million hours) in April 2026. Each is roughly 10x the previous release.
Can Egocentric-1M be used commercially?
Yes. It is Apache 2.0 licensed, permitting commercial use, modification, and redistribution. Access requires accepting the dataset conditions on Hugging Face first.
What domains does Egocentric-1M cover?
Real production environments — assembly lines, sorting, packaging, and machining — recorded from the first-person perspective of factory workers wearing head-mounted cameras. This industrial focus distinguishes it from home-focused datasets like Ego4D.
What does 'the internet for physical AI' mean?
Build AI founder Eddy Xu used this phrase to position Egocentric-1M as the foundation-scale pretraining corpus for embodied AI — playing the role for robot learning that web-scale text corpora played for large language models.