EXCERPT / 01
Remove tape
Removing pieces of adhesive tape from an air conditioner’s front housing.
28.98 s870 frames
Inside the release · RGB · Human observation
Hands-on maintenance. A closer point of view.
Chest-mounted iPhone footage of hands-on appliance maintenance. Explore six RGB excerpts, inspect the optional hand and object estimates, and review the approximate action intervals.
Download on Hugging FacePublic release · CC BY 4.0 · 2026-09-08
6
Curated excerpts
176.84 s
Total duration
5,308
RGB frames
RGB
Observation data
Inside the session
Six excerpts. Everyday tool use.
EXCERPT / 02
Brush front grille
Brushing the front grille while stabilizing the housing.
29.95 s899 frames
EXCERPT / 03
Brush housing
Preparing the brush and scrubbing the exposed upper housing.
29.98 s900 frames
EXCERPT / 04
Brush panel channels
Holding the detached panel and brushing its narrow channels.
29.98 s900 frames
EXCERPT / 05
Wipe panel
Wiping panel surfaces while stabilizing the panel with the other hand.
29.98 s900 frames
EXCERPT / 06
Wipe panel corners
Wiping recessed corners and channels with a paper towel.
27.95 s839 frames
The release covers one wearer, one appliance/work area and one apparent session. These are six excerpts, not six independent trials. They do not document a verified completed maintenance procedure or working appliance.
Understand each channel
Observation, with its context intact.
Public RGB
Traceable video excerpts
1920 × 1080, approximately 30 fps. The public 8-bit SDR/H.264 clips are tone-mapped from iPhone 15 Pro HLG HDR recordings, with original presentation-time spacing and one-to-one source-frame mapping. Full original MOV files remain with the owner.
Estimated & curated
Hands, objects & actions
Optional hand and object annotations are model estimates. Thirteen action intervals were prepared by an AI assistant from visual review; boundaries are approximate to two seconds and have not been independently checked by a human annotator.
Unavailable
Metric geometry & control
Depth, metric camera calibration/pose, IMU and measured robot commands are unavailable. There are no execution-success labels, force or tactile measurements. Audio and source-container metadata are omitted from public video.
Evaluate before training
What to inspect first.
- Load an excerpt and compare episode timestamps with source timestamps and frame mappings. Expect depth and metric 3D fields to be unavailable.
- Inspect tool visibility, bimanual occlusion and repetitive cleaning. Review action boundaries against video before treating them as evaluation labels.
- Compare raw and temporal hand estimates. Temporal processing can briefly retain three tracks; track IDs are not verified physical identities.
- Keep this session together when forming splits. Use independently reviewed labels and held-out sessions for claims about generalization.
Coverage is not accuracy.
Unknown handedness ranges from 13.7% to 100% across excerpts. Hand estimates were requested at 10 Hz and objects at 5 Hz; emitted timestamps do not measure recall. There is no independent annotation-accuracy result or demonstrated training gain.
Inside the download
- videos/
- Public SDR training excerpts. These are processed derivatives, not the original HDR MOV files.
- annotations/
- Raw and temporal hand/object estimates, action intervals and source-frame mappings.
- review/
- Overlay videos for qualitative inspection; their display clock is not the alignment reference.
- quality/
- Per-excerpt channel coverage and missingness.
- source_index.json
- Selection windows and authorization record.
- SCHEMA.md
- Clocks, transforms, units and missing-data behavior.
From sample to pilot
Test a question this footage can answer.
Use this release to begin evaluating video-language alignment, temporal action understanding or hand/object interaction workflows. A useful initial exercise is to distinguish brushing from wiping during occlusion, with your own reviewed reference labels. That is a proposed evaluation, not a reported result.
For a broader collection, specify complete task sequences, multiple workers and sites, tool visibility, camera placement and acceptance criteria. Our dataset brief template helps turn those decisions into a scoped pilot.
Try one episode.
Use the portable reader included with the release. These commands select the documented revision so you can reproduce this page’s reference package.
pip install huggingface_hub
hf download diffracting/egocentric-maintenance-sample --repo-type dataset --revision acd7da07ca85bf16c99c61edf656515fa9d2b9c4 --local-dir diffraction-sample
cd diffraction-sample
pip install -r requirements.txt
python observation_reader.py .See the Hugging Face dataset card for the latest files and usage examples. Smaller browser previews are distinct from the training videos in the download.
License and attribution
This release is labeled CC BY 4.0. Follow the license terms and attribute Diffraction Egocentric Maintenance Sample (2026), the dataset URL and the revision used. This page summarizes revision acd7da07ca85; consult its documentation for the full schema and limitations.
Continue exploring
A different task. A different signal.
Take the next step
Need a collection around your task?
Use what you learn from the sample to specify the activities, signals and acceptance checks your own project needs.
Build your brief with the buyer’s guide →