Egocentric-10K

Build AI's open-source dataset of roughly ten thousand hours of first-person video recorded by factory workers.

Egocentric-10K is a first-person video dataset released on Hugging Face in 2025 by the U.S. startup Build AI, under the Apache 2.0 license. It was recorded by workers wearing monocular head-mounted cameras while working in 85 real factories, totaling about 10,000 hours across 192,900 video clips and 1.08 billion frames, at 1080p and 30 frames per second, with fisheye camera intrinsics included. It emphasizes footage where the hands stay visible and are actively manipulating things, and is positioned specifically as data for robot learning rather than for general video understanding. The dataset carries no action labels, so it's typically used for human-video pretraining or latent-action pretraining: a model first learns visual representations and action priors from large amounts of human hand manipulation, then gets fine-tuned on a small amount of robot data. According to its website, Build AI has since released larger follow-up versions, Egocentric-100K and Egocentric-1M.

ExampleA typical clip in the dataset shows a worker wearing a head-mounted camera at their station, fitting a part into a jig with both hands visible throughout — well suited to learning hand manipulation.

Related

Egocentric Video · Human Video Data · Action-free Video · Pretraining on Human Videos · Latent Action Pretraining · In-the-wild Data

Sources

Egocentric-10K (Hugging Face dataset card)
Build AI 官网 (Chinese)

As of

2026-03