EPISODE / 01
Onion preparation
Preparing and peeling an onion on a cutting board.
75.73 s4,543 frames
Inside the release / 01 · Human observation
An egocentric kitchen dataset. Open to inspection.
Three everyday manipulation recordings captured with Stray Scanner on a LiDAR-equipped iPhone. Explore the native RGB-D, inspect the estimates, and understand the limits before using them.
Download on Hugging FacePublic release · CC BY 4.0 · September 8, 2026
3
Recordings
167.85 s
Total duration
10,070
Native RGB frames
RGB-D
Primary capture
The capture sequence
Three tasks. One small collection.
EPISODE / 02
Pepper slicing
Preparing and slicing a bell pepper with a kitchen knife.
57.75 s3,465 frames
EPISODE / 03
Pepper dicing
Dicing bell pepper strips on a cutting board.
34.37 s2,062 frames
Participant count, exact device model, app version and independent session identity were not supplied in the release. These closely related kitchen recordings do not establish cross-person or cross-environment generalization.
Read the evidence correctly
What’s captured. What’s inferred.
Native capture
RGB, depth & device sensors
Native 1920 × 1440 RGB video, 256 × 192 metric depth and confidence, per-frame calibration and raw device IMU. IMU acceleration units require capture-version verification; device motion is distinct from wrist motion.
Estimated channels
Pose, hands & objects
Camera pose comes from device visual-inertial odometry. Hand pose, palm orientation, object tracks and contact are estimates. Missingness and identity ambiguity are documented; confidence scores are not calibrated error probabilities.
Outside this release
Robot control & performance
No measured robot commands, force or tactile data. No independent annotation accuracy benchmark or demonstrated robot-policy improvement. The recordings share one conservative session group; there is no train/test benchmark.
Begin with a technical evaluation
A useful first pass.
- Load a native episode and inspect corresponding RGB, depth, calibration and missing values.
- Compare optional annotation overlays with the video, especially fast cutting and hand occlusion.
- Read the quality report before choosing which fields your model can consume.
- Define a separately labeled evaluation and your own downstream experiment before drawing performance conclusions.
Packaged validation checks cover reader behavior, geometry, file hashes and missing-value semantics. They establish packaging integrity, not label accuracy. The sample’s low-resolution depth can miss fingers, thin tool edges and occlusion boundaries.
Read the full quality report ↗Inside the download
- videos/
- Native RGB recordings; use these for original training pixels.
- sensors/
- Depth, confidence and source sensor CSVs.
- annotations/
- Standalone Parquet tables for derived annotations.
- review/
- Annotated overlays for visual inspection.
- quality/
- Per-episode coverage, missingness and provenance reports.
- SCHEMA.md
- Units, frames, clock mapping and missing-data behavior.
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-kitchen-sample --repo-type dataset --revision 75ccd3fc9fbb78be56029539ab1ebf9fb1f3e2f4 --local-dir diffraction-sample
cd diffraction-sample
pip install -r requirements.txt
python observation_reader.py .For the latest files, usage examples and requirements, visit the Hugging Face dataset card. The browser viewer uses smaller previews; the download includes native video.
License and attribution
The release is labeled CC BY 4.0. Follow its terms and the dataset card’s attribution request: “Diffraction, Egocentric Kitchen Capture Sample (2026),” the dataset URL and the revision used. This page summarizes the public release at revision 75ccd3fc9fbb; consult its source documentation for the full schema and limitations.
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 →