Duet Request a sample

A new data layer for physical AI.

Two people, two stereo headsets, one shared frame. Duet records paired egocentric video of humans working together and reconstructs camera poses, hands, scene geometry, and the steps and handovers between them.

demo recording s017
184.0 s · 14 steps · 5 handovers
dorm kitchen · 2 × ZED Mini
drag to orbit · meters · demo data
ego A
ego_A.mp4 · faces blurred
ego B
ego_B.mp4 · faces blurred
00:00.0
03:04.0
tracking confidence, per frame
1.0
0.2
Fig. 1 — Shared world frame, meters. Camera trajectories and 21-point hands for A (clay) and B (lake); table plane fitted from scene geometry; diamonds mark handovers. Below: per-frame hand-tracking confidence, dips are shown, not smoothed.
s017 · dorm kitchen · 2 × ZED Mini @ 30 fps · 5,520 frames · sync offset ±4 ms · pipeline v0.7 · exported 2026-10-06 · demo data
What is in the data
synchronized stereo video
2 × 1280 × 720, 30 fps, ±4 ms
camera poses
6-DoF per frame, meters, shared frame
hand keypoints
21 per hand, 3D, with confidence
scene geometry
Table plane, depth, static layout
step & handover labels
Half-open intervals, primitive, giver, receiver
speech
Timestamped transcript per person
How we collect it

Workplaces get a work-efficiency analysis. Every session they receive becomes labeled training data.

Record a shift
two headsets, one task
→
Deliver the analysis
timeline, waste, SOP
→
Grow the dataset
labels paid for by the pilot
Scale
100 h
v1 · dorm kitchen, two cameras
now
1,000 h
v2 · factories and workplaces, custom stereo headsets
next
10,000 h
and up, across sites and tasks
then
01
Shared world

Both videos and the 3D frame on one scrub bar. Toggle trajectories, hands, table, labels.

Open →
02
Interaction timeline

Steps per person, handovers between lanes, coordination events. Click a segment to seek.

Open →
03
Workplace analysis

The same task across three sessions: time split by lean waste, step durations, findings with evidence.

Open →
04 Sixteen 10 s clips at 6× with hand and body overlays, plus the 3 min compilation.
Demo clips →
Request a sample of the dataset: research@duetlabs.co
Demo data. Confidence shown per frame; failed tracking is shown as a dip, not hidden. Units: seconds, meters.
Duet Request a sample

A new data layer for physical AI.

Two people, two stereo headsets, one shared frame. Paired egocentric recordings of humans working together, reconstructed and labeled by step and handover.

demo recording s017 · 184.0 s
14 steps · 5 handovers · dorm kitchen
t 00:47.3
A 0.93  B 0.88 conf
still · open on desktop to scrub
Fig. 1 — Shared world frame, meters. A (clay), B (lake), fitted table plane.
ego A
ego B
What is in the data
synchronized stereo video
camera poses
hand keypoints
scene geometry
step & handover labels
speech
How we collect it

Workplaces get a work-efficiency analysis. Every session they receive becomes labeled training data.

Scale
100 h
v1 · kitchen
1,000 h
v2 · factories
10,000 h
and up
01 Shared world
02 Interaction timeline
03 Workplace analysis
Interactive views open on a laptop or larger. This page is read-only.