Lightwheel and Hugging Face release 100k-hour egocentric dataset for Physical AI

vanstriendaniel · x · 2026-08-27

Lightwheel AI, in partnership with Hugging Face, has released EgoSuite-Open100K, an open-source egocentric human activity dataset designed for Physical AI. The dataset aims to total 100,000 hours across 15,000+ tasks and real-world scenes, with the first 10,000 hours now available. It provides first-person video demonstrations of human actions—reaching, grasping, sequencing, and recovery—offering scalable supervision signals that robot-only data collection struggles to match. The data supports LeRobot v3 and MCAP formats for both academic and commercial use. Citing NVIDIA's EgoScale work, the release highlights the proven log-linear scaling potential of egocentric data.

Original post →

More from Embodied

Embodied channel →