Prototype concept · July 2026
Robots learn from
human hands.
WorkTrace turns real physical work into the training data that teaches machines to carry, assemble, clean, and correct. Captured from a hat brim. Labeled in the flow of work.
The market
The bottleneck isn't compute.
It's hands.
Robot foundation models are starved for one input: diverse, labeled demonstrations of skilled hands working in real places. Labs already buy it by the hour.
The system
Capture in the
flow of work.
Wear
A small camera in a hat brim points at the hands. The worker just works.
Track
21 landmarks per hand, tracked live, with a hands-visible meter that catches bad framing before it costs an hour.
Label
Task chips segment the episode as it happens. Quality is scored before anyone leaves the site.
The wedge
Start where
consent is clean.
Home staging
Vacant houses. Furniture carries, assembly, styling. A crew of two produces synchronized multi-view episodes of every task.
Rental turnover
Empty short-term rentals on a schedule. The same tasks, hundreds of times, across hundreds of rooms.
Hotel floors
Standardized rooms worldwide. A floor in Manila trains the same robot as a floor in Chicago.
Offshore arms researched and ranked: the Philippines first, Colombia second. Occupied homes come last, once the privacy pipeline has earned it.
The product
Annotations and rights.
Not raw footage.
- World-frame camera pose from the rig's own telemetry
- 3D hand pose with per-joint confidence
- Language-labeled task segments
- A signed consent record inside every episode manifest
- Delivered in the formats labs actually train on
Raw video is nearly free. Provenance is not.
The film
Sixty seconds,
start to finish.
See it
working.
The capture app, the hardware spec, and the business research are all real. Plug in a webcam and watch your own hands become data.