license: cc-by-4.0
pretty_name: Diffraction Egocentric Maintenance Sample
configs:
- config_name: default
data_files:
- split: sample
path: data/episodes.parquet
language:
- en
tags:
- video
- egocentric
- human-demonstration
- robotics
- physical-ai
- maintenance
size_categories:
- n<1K
Diffraction Egocentric Maintenance Sample
Chest-mounted iPhone video of hands-on appliance maintenance: tape removal, brushing, panel handling and wiping recessed surfaces. 6 curated excerpts complement Diffraction's RGB-D kitchen sample with a different task domain.
This is human RGB observation data for evaluating video-language, temporal action understanding and hand/object interaction workflows. Depth, metric camera calibration/pose, IMU, robot commands and execution-success labels are unavailable. Optional hand/object annotations are model estimates with no independently measured accuracy.
| Excerpt | Seconds | RGB frames | Instruction |
|---|---|---|---|
| remove-tape | 28.98 | 870 | Remove pieces of adhesive tape from the air conditioner front housing. |
| brush-front-grille | 29.95 | 899 | Brush the air conditioner front grille while stabilizing the housing. |
| brush-housing | 29.98 | 900 | Prepare the brush and scrub the exposed upper housing of the air conditioner. |
| brush-panel-channels | 29.98 | 900 | Hold the detached front panel and brush along its narrow channels. |
| wipe-panel | 29.98 | 900 | Wipe the detached panel surfaces while stabilizing the panel with the other hand. |
| wipe-panel-corners | 27.95 | 839 | Wipe the recessed corners and channels of the detached panel with a paper towel. |
176.84 seconds · 5,308 RGB frames · one session · 13 approximate action intervals.
Download and use
pip install huggingface_hub
hf download diffracting/egocentric-maintenance-sample --repo-type dataset --local-dir diffraction-maintenance
cd diffraction-maintenance
pip install -r requirements.txt
python observation_reader.py .
from observation_reader import ObservationDataset
sample = ObservationDataset('.')
key = next(iter(sample.episodes))
frame = sample.frame(key, 100)
print(frame['rgb'].shape, frame['timestamp_s'], frame['source_timestamp_s'])
assert frame['depth_m'] is None
hands = sample.annotations_near(key, 'hand_pose', frame['timestamp_s'], max_age_s=0.05)
Capture and processing
The source files identify an iPhone 15 Pro recording 1920 x 1080 HLG HDR video at approximately 30 fps. The owner described a chest-mounted camera setup and authorized the public sample. Full original MOV files are preserved locally with SHA-256 provenance. Public RGB clips are documented 8-bit SDR/H.264 excerpts with original presentation-time spacing and one-to-one source frame mapping. Audio and source-container metadata are omitted. Smaller browser videos are previews; use videos/ for the training representation.
All excerpts appear to belong to the same maintenance session. They are one technical sample, not six independent trials or a train/test benchmark. No claim of a completed maintenance procedure or verified appliance operation is made.
Quality and intended use
QUALITY.md reports review coverage and known failure modes; validation.json records the executable release checks. SCHEMA.md defines clocks, transforms, units and missing-data behavior. Per-excerpt quality/ reports provide coverage and missingness; review overlays expose optional estimated annotations. Occlusion, close tools, deformable paper and repetitive cleaning are difficult cases. A detection score is not a calibrated error probability. Handedness conflicts become unknown. Any language intervals are explicitly curator annotations with a stated review method.
A downstream lab should evaluate relevant labels and model outcomes before scaling training. This sample demonstrates capture and delivery structure; no robot-policy performance gain or independent annotation-accuracy result is claimed.
Verify and evaluate
pip install -r requirements-validation.txt
# Install FFmpeg/ffprobe on PATH, then:
python validate_release.py .
checksums.sha256 covers every release artifact. processing_record.json records processing tools, source/derivative lineage and the distinction between batch execution and final assembly code. source_index.json records selection windows and the owner authorization record. The public preview catalog can also be loaded with datasets.load_dataset("diffracting/egocentric-maintenance-sample", split="sample"); install the video decoder dependencies appropriate to your environment, or use datasets.Video(decode=False) to retrieve embedded preview bytes.
Work with Diffraction
This sample is a starting point for task-specific professional egocentric capture. A pilot can define the activities, workplaces, camera setup, review protocol and measurable delivery criteria with the receiving lab. Contact Diffraction through the organization page or start a dataset discussion.
License and attribution
CC BY 4.0 permits reuse under its attribution terms. Attribute Diffraction, Egocentric Maintenance Sample (2026), this dataset URL and the revision used. See LICENSE.