Robot learning dataset
EgoDex
Egocentric human manipulation recordings collected with Apple Vision Pro, combining video, tracked head/body/hand poses and language annotations across 194 tabletop tasks. This is human demonstration data rather than humanoid-robot recordings, intended for learning and evaluating dexterous manipulation.
Records last verified Oct 13, 2026 · Send a correction
Licence & use
CC-by-NC-ND
Non-commercial use
Licence text ↗Terms apply to the checked release. Read publisher conditions for mixed collections, converted formats and gated access.
- Publisher
- Apple
- Released
- Not dated
- Published size
- 829 h
- Modalities
- rgb, human-pose, language
- Formats
- HDF5, MP4
- Real / simulated
- Real
- Includes humanoid recordings
- No
Size & release notes
Official README and current paper publish 829 hours at 30 Hz across 194 tasks. Download parts are individually listed as 300 GB, 16 GB and 200 GB; no summed storage or TB conversion is stored. README states CC-by-NC-ND without a version. Its LICENSE file governs software only, so no dataset-license version is inferred.
More robot learning datasets
Evidence
Sources
- [s1] EgoDex official dataset README: formats, downloads and dataset terms ↗
Apple · official · accessed Oct 13, 2026
- [s2] EgoDex software license (not the dataset license) ↗
Apple · official · accessed Oct 13, 2026
- [s3] EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video ↗
arXiv · paper · published May 16, 2025 · accessed Oct 13, 2026