Hub drops HubEgo-100K egocentric robotics dataset
Hub open-sourced HubEgo-100K on Hugging Face under CC BY 4.0, releasing an initial tranche of 10,124 hours of first-person RGB and IMU recordings from 12,000+ contributors across real-world workplaces. Packaged in LeRobot and MCAP formats, the dataset provides unscripted demonstrations across kitchens, workshops, and construction sites to train vision-language-action and robotics foundation models.
The robotics data bottleneck has long been artificial: training million-dollar humanoids on sterile lab manipulation fails the moment machines encounter unstructured real-world clutter. HubEgo-100K delivers the massive egocentric diversity embodied AI desperately needs, packaged directly into Hugging Face's LeRobot ecosystem.
- –Distribution shift fix: Moving beyond pristine lab environments exposes policies to noisy backgrounds, varied lighting, and real-world tools.
- –Native LeRobot compatibility: Shipping over 72,000 episodes in LeRobot v3 and MCAP formats eliminates data-wrangling friction for teams training VLA policies.
- –Multimodal motion tracking: Synchronized head-mounted 1080p video and IMU data captures the biomechanics of real human dexterous manipulation.
- –Commercially permissive: The CC BY 4.0 license lowers entry barriers for robotics startups priced out of proprietary teleoperation data networks.
- –Scalable physical AI pretraining: This 10,000-hour initial release toward a 100,000-hour roadmap provides embodied AI with the scale of pretraining video that language models enjoyed years ago.
DISCOVERED
47m ago
2026-11-16
PUBLISHED
1h ago
2026-11-16
RELEVANCE
AUTHOR
LinusEkenstam