Robot-learning datasets often trade scale for action consistency: large video collections rarely share a common representation of hand motion. This release addresses that mismatch by curating data from five sources and aligning camera coordinates, wrist axes, and hand or gripper openness, making human video more usable for embodied pretraining.
What Sets It Apart
- The corpus contains 50,000 hours across household, commercial-service, and industrial manipulation, giving researchers varied human demonstrations rather than robot-only trajectories.
- It combines HumanNet, EgoSuite-Open100K, EgoDex, Gen-HumanEgo, and EgoVerse. The report counts 35.9 million selected clips, while the released data contains 31,396,799 physical episodes; these are different counting units.
- Hand visibility, motion stability, and manipulation activity guide curation. Language and hand-motion annotations pair task context with action signals, supporting research into how human demonstrations can inform robot policies.
- The release uses LeRobot v3, with video, action/state data, metadata, and provenance. Subset indices cover selections from 100 to 50,000 hours, allowing experiments at different data scales.
Who It’s For—and the Trade-off
A good fit for researchers studying egocentric pretraining, data scaling, and transfer from human manipulation to robots. It is less suitable for users seeking a uniformly licensed, ready-to-train single-source dataset: source-specific terms apply, and dataset size alone does not guarantee consistent annotation quality or downstream performance.
