PHYSICAL AI · DATA INFRASTRUCTURE
Real-world data for robots, starting with the trades
We capture and curate consented, contact-rich manipulation data from electricians, HVAC techs and other skilled tradespeople
THE BOTTLENECK
Robots don't have an ImageNet
Foundation models for robotics are constrained by data, not compute. The hardest data to get, real humans doing dexterous physical work in messy environments, barely exists. We build exactly that, at the source
$18.8B
invested in robotics in H1 2026
350 hrs
DROID, one of the largest open manipulation datasets, took 18 labs a year to collect
#1
data is the leading constraint on model progress
Industry context, not company metrics
WHAT WE DO
Source at the frontline
Real tradespeople, real jobs, captured on the actual work with head and wrist mounted cameras
Curate for training
We turn raw footage into verified, annotated, coverage-designed episodes, so you buy accepted units, not raw hours
Consented & rights-cleared
Every hour is captured under explicit consent with clean provenance, built for EU privacy standards
HOW IT WORKS
01 · You define the tasks
Tell us the manipulations and edge cases your models are failing on
02 · We capture & curate
Our trade network records them; we gate for quality, privacy and coverage
03 · You get training-ready data
Synchronized RGB-D, IMU and world-frame 6-DoF pose, with 3D hand poses and per-segment action labels, exported in the formats teams already train on such as LeRobot and HDF5
On the roadmap: force and tactile capture, so buyers get the contact signals that video alone can never recover
Building physical AI?
Run a trade business? Get paid for footage of your normal jobs.
Learn more