Densely labeled egocentric recordings of real human work, captured across homes and workplaces.

What is in every frame

We record and annotate our own footage, so a frame carries more than pixels.


RGB	first person video of real work, not staged demonstrations
Metric depth	iPhone Pro LiDAR, in metres, not relative
Camera pose	6-DoF per frame, from ARKit tracking
Hand pose	MANO joints for both hands
Contact	per hand and per object, with a validity flag on every value
Segmentation	hand and object masks
Language	dense sub step labels, time aligned

Everything ships as LeRobot v3.0, so it loads in one line with no conversion step.

Datasets

everyday-manipulation-3d: the labeled corpus. 32 episodes packaged as LeRobot v3.0, every frame carrying metric depth, camera pose, MANO hand pose, per hand object contact, masks and time aligned language. CC BY NC SA 4.0, research use.
everyday-manipulation-3d-raw: 1,513 clips of everyday two handed manipulation, 10.28 hours, with synchronised RGB, metric LiDAR depth, ARKit camera pose and per frame intrinsics. CC BY 4.0, commercial use allowed with attribution.
caryx-viewer-assets: browser streaming previews behind the viewer at caryx.ai/data.

The annotated release, adding hand pose, contact and segmentation on top of these clips, is in review. More collections are in production.

Custom collections, larger corpora, and commercial licensing: founders@caryx.ai