Dataset · Apple
EgoDex
Download from ml-site.cdn-apple.com/datasets/egodex/ (Parts 1-5 at 300 GB each, test 16 GB, extra 200 GB). 3D hand/finger/body pose annotations included (not one of the fixed modality tags).
Facts
- Organisation
- Apple
- Form tags
- human hand · human video
- Embodiments (as stated)
- human hands (egocentric video, Apple Vision Pro)
- Episodes
- 338,000
- Hours
- 829
- Frames
- unverified
- Tasks
- 194
- Scenes
- unverified
- Modalities
- rgb · language
- Format
- HDF5 + MP4 (paired per episode)
- Licence
- CC-by-NC-ND terms
- Commercial use
- not allowed (stated)
- Released
- 2025-05
- Size
- 1500 GB
Sources · 14 sources
name https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Abstract: 'we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation'
org https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Abstract: 'we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation'
url https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Abstract: 'we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation'
embodiments https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Abstract: 'we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation'
form_tags https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Abstract: 'we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation'
episodes https://arxiv.org/html/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet Paper body: '338000 task demonstrations'
hours https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet '829 hours of egocentric video ... 194 different tabletop tasks'
tasks https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet '829 hours of egocentric video ... 194 different tabletop tasks'
modalities https://github.com/apple/ml-egodex Official page · Fetched 2026-09-18 · No snapshot stored yet '829 hours of 30 Hz 1080p egocentric video'; 3D pose annotations for head, upper body and hands; natural language annotation; camera intrinsics
format https://github.com/apple/ml-egodex Official page · Fetched 2026-09-18 · No snapshot stored yet HDF5 files paired with MP4 files; 'approximately 1.5 TB' total (train ~725h, test ~7h, extra ~97h)
size_gb https://github.com/apple/ml-egodex Official page · Fetched 2026-09-18 · No snapshot stored yet HDF5 files paired with MP4 files; 'approximately 1.5 TB' total (train ~725h, test ~7h, extra ~97h)
license https://github.com/apple/ml-egodex Official page · Fetched 2026-09-18 · No snapshot stored yet README: released under 'CC-by-NC-ND terms'
commercial_ok https://github.com/apple/ml-egodex Official page · Fetched 2026-09-18 · No snapshot stored yet README: released under 'CC-by-NC-ND terms'
release_date https://arxiv.org/abs/2505.11709 Official page · Fetched 2026-09-18 · No snapshot stored yet arXiv v1 2025-05-16 (v3 2026-03-09)
Related bodies in the index
Matched by stated robot names or form; a match means the form is covered, not that a model was trained on it.
—
Clear direction, grounded judgement.