FULL STORY
HubEgo-100K: From Teaser to Open Release
Hub teased a partnership with Hugging Face's LeRobot team to open-source the world's most diverse egocentric robot dataset, then released HubEgo-100K on Hugging Face with an initial 10,000 hours of first-person skill videos.
2026-11-11 ~ 2026-11-17 · 2 episodes · 7 posts
Episode 1 · Hub to Open-Source World's Most Diverse Egocentric Robot Dataset (2026-11-11, 2 posts)
Robotics startup Hub will release the world's most diverse egocentric dataset on Hugging Face on November 12 in partnership with LeRobot. The company spent two years paying people to film their daily work in first person, aiming to close the data gap between robots and LLMs.
- Hub to open-source most diverse egocentric robot dataset on Hugging Face Nov 12 — alex_verem · 2026-11-11
- Robots Lack an Internet to Learn From: Hub to Open-Source Largest Egocentric Dataset — LinusEkenstam · 2026-11-11
Episode 2 · Hub open-sources HubEgo-100K, a 100K-hour egocentric skill video dataset (2026-11-17, 5 posts)
Hub, together with the Hugging Face LeRobot team, has open-sourced HubEgo-100K on Hugging Face—billed as the most diverse egocentric dataset to date, released under CC BY 4.0 and free for research and embodied AI training. Several authors see the move as breaking the monopoly on robotics learning data.
Confirmed
- The first release (release 1) is live at the 10,000-hour scale: m1/m4 give the exact figures of 10,124 hours, 44.7TB, and about 1.09 billion frames, totaling 72,754 clips with a median length of 4.9 minutes and a maximum of 30 minutes; m3 notes this batch was recorded by 12,227 contributors over 72 hours
- The data consists of first-person footage of real skill tasks, covering 15,548 real environments, 56 task types, and 30+ countries/regions (m3), with scenes including construction sites, workshops, salons, kitchens, stores, warehouses, and homes (m5)
- The dataset comes from 100,000+ contributors, targets 100,000 hours in total, and will be expanded in stages over the coming weeks
- Licensed under CC BY 4.0, free for research (m3/m5)
Why it matters
- First-person videos of skilled manual work are scarce training material for robot and embodied AI skill learning, and large-scale data of this kind has largely been monopolized by a few institutions; @lukasmziegler explicitly called it "the robotics learning data monopoly being broken"
- The dataset's emphasis on real environments and task diversity (56 task types, 15,000+ environments) is directly valuable for generalization research
- 100,000 hours of egocentric skill data open-sourced on Hugging Face, free for all — lukas_m_ziegler · 2026-11-17
- HubEgo-100K: 100K hours of egocentric skilled manual work video released — rohanpaul_ai · 2026-11-17
- 10K-hour egocentric video dataset released: 72,754 head-mounted episodes for robot learning — rohanpaul_ai · 2026-11-17
- HubEgo-100K: 12,227 contributors' egocentric work videos, open under CC BY 4.0 for robot learning — rohanpaul_ai · 2026-11-17
- Hub open-sources HubEgo-100K: 44.7TB egocentric video dataset for robots — rohanpaul_ai · 2026-11-17