EgoVerse — Egocentric Robot Training Dataset: 8.5M Frames, 74 Countries

What is it?

EgoVerse is the largest open-source egocentric robot learning dataset available as of 2025. Developed by Georgia Tech RL2 in collaboration with Mecka AI, ETH Zurich, and Scale AI, it contains 57,761 labelled episodes totalling 1,213 hours of real-robot demonstration data. The dataset covers 5,059 distinct tasks across six robot embodiments — bimanual arm platforms from Aria, Eva, Mecka, and Scale — collected in diverse real-world environments including kitchens, bedrooms, dining rooms, offices, workshops, and outdoor balconies. Data is distributed in Zarr and LeRobot format via AWS S3, with an interactive episode browser at partners.mecka.ai/egoverse.

Who is it for?

EgoVerse is designed for machine learning researchers working on imitation learning, behaviour cloning, and vision-language-action model training for household and service robot applications. It is particularly valuable for researchers who need large-scale, diverse, commercially licensable manipulation data across multiple robot platforms. The dataset is also a primary reference for policy researchers studying generalisation across embodiments, as it covers six distinct bimanual arm configurations under a single consistent task taxonomy.

Key specifications

• •Episodes: 57,761 labelled demonstrations
• •Duration: 1,213 hours of recorded data
• •Embodiments: aria_bimanual, aria_left_arm, aria_right_arm, eva_bimanual, eva_right_arm, mecka_bimanual, scale_bimanual, scale_left_arm, scale_right_arm
• •Tasks: 5,059 distinct tasks across cleaning, dishwashing, folding, cooking, manipulation, pick-and-place, and human-robot-interaction categories
• •Format: Zarr, LeRobot (HDF5)
• •License: CC BY-SA 4.0 — commercial use permitted with attribution
• •Access: Open — freely downloadable from AWS S3

How it compares

EgoVerse is the largest open manipulation dataset by episode count and total hours. Open X-Embodiment covers more embodiments (22 vs 6) but has fewer hours of egocentric household data. DROID (76,000 episodes) is comparable in scale but focuses on Franka Panda arm manipulation in lab settings rather than household environments. EgoVerse's egocentric camera perspective and diversity of real-world home environments make it the strongest dataset for training domestic service robot policies.

Limitations and access notes

EgoVerse covers manipulation tasks but does not include locomotion, outdoor navigation, or physical contact-rich tasks requiring force feedback. The dataset was collected using bimanual arm platforms — it does not include mobile base data, humanoid full-body data, or legged robot data. The CC BY-SA 4.0 license requires derivative datasets to be shared under the same license. Access is free and requires no registration.