See the data. Then put it to work.
Follow a real battery-change episode from first-person video to timed actions, then load the same files in Hugging Face Datasets or FiftyOne.
The action sequence
Click an action to seek. Overlaps and gaps are retained.TaskChange a GoPro battery
TaxonomyTask → action → verb + noun
Source mappingClip time + 29s = source time
Original labels are displayed as supplied; “battery_door” is shown with a space for readability. Source correctness judgments are inherited. No hand tracking, depth or robot actions are included. Audio is removed.
Source recording and annotations: HoloAssist. Handtrained adds the excerpt, packaging and import checks. CDLA-Permissive-2.0 license.
Download the complete sampleA sample you can load.
The v1.0.1 package contains an MP4, temporal action annotations, a manifest, provenance, checksums, a Hugging Face Datasets notebook and a native FiftyOne export. Local import is tested on macOS arm64 with Python 3.12; Colab is not yet verified. The package is not published to the Hugging Face Hub.
Read the complete loading instructions · Download the notebook
What this sample
lets you evaluate.
Inspect a complete manipulation sequence, trace a label back to its original annotation and run the supplied loaders. The release includes checksums and a validation report.
A successful import is one part of acceptance. Task fit, capture quality, label accuracy and usage rights still need buyer-specific review.
- One 38-second video, 896 × 504 pixels
- 16 fine actions and one coarse task annotation
- Original IDs, source times and action attributes
- 1,140 frames at a normalized 30 fps
- VideoFolder metadata and native FiftyOne labels
- Source license, provenance and pinned setup
30 fps is a delivery conversion; it adds no captured information. Natural motion blur, hand occlusion, label overlaps and unannotated moments remain visible.
Discuss your
data needs.
Bring a brief. Or start with a task. We’ll define the next step together.