Duet

Duet: Human-human interaction data for physical AI

Request a sample

Loading the 3D view…

Head cameras · 8× speedA's handsB's hands
A · Head camera
B · Head camera
Session 0:00.0 / -:--.- 0 / 0

Steps, handovers and coordination are labeled automatically by a vision-language model from both head cameras every 2 s.

Working togetherWorking independentlyWaiting or askingNon-value-add stepHandover (giver's color)

SpaceK play · JL ±5 s · ←→ 1 s (Shift 10 s) · ↑↓ handovers · <> speed · +−0 zoomDrag the ruler to scrub · swipe to move · tap a step to jump · pinch to zoom

01 / Capture

Capture spec

Shared worldOne world frame for both head cameras and the wearers' hands, in meters, z up.
02 / Roadmap

Hours of data

1 dot = 10 h
v1 · target100 hDorm. Two head cameras.
v2 · planned1,000 hFactories and workplaces. Custom stereo headsets.
v3 · planned10,000+ hMany sites and tasks.