EgoVerse

A continually growing platform of first-person human demonstration data for robot learning, built jointly by universities and companies.

EgoVerse is an open dataset and platform led by Danfei Xu's group at Georgia Tech, together with Stanford, UC San Diego, ETH Zurich, and other universities, plus companies including Meta, Scale AI, and LightWheel AI (光轮智能); its paper was released in April 2026. The version described in the paper contains 1,362 hours and about 80,000 human demonstrations, covering 1,965 tasks, 240 scenes, and 2,087 demonstrators; the project's website describes it as a continually growing dataset, with the current version at roughly 4,000 hours. The data uses a unified format, with camera pose, 3D head tracking, and dense language annotations included. The paper runs human-to-robot transfer experiments with a shared pipeline across multiple labs and robot types, and concludes that more human data generally improves the resulting policy — but only when the human data actually matches the task the robot needs to learn.

ExampleIn the paper's “place cup on saucer” bimanual task, multiple labs each running their own robot use the same pipeline to compare success rates as different amounts of human data are added.

Also called

EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World

Related

EgoMimic · Egocentric Video · Human Video Data · Cross-Embodiment Data · Data Scaling Laws in Imitation Learning (Robotic Manipulation) · Crowdsourced Data Collection

Sources

EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World (arXiv)
EgoVerse 官网 (Chinese)

As of

2026-09