DDLS · egocentric trades data

Training data from the trades, seen through their eyes.

Head-camera recordings of electricians, plumbers and carpenters on real jobs, turned into hand, object and task data for robot learning, and checked by replaying every session in physics.

See the data

Engraving of a plumber kneeling at a kitchen cabinet fitting a pipe, while a robot arm on a bench makes the same grip, dotted lines joining their joints.

Pl. IThe plumber and the robot, the same grip.

Egocentric

Seen from the worker's own eyes.

A small camera in the bridge of the safety glasses tradespeople already wear. It sees what they see, at the distance and angle a robot's own camera will: their hands, their tools and the work in front of them.

What’s in a dataset

Engraving of the view through a pair of safety glasses: the wearer's own hands stripping a wire at an outlet box, dots on each knuckle.

Real job sites: Kitchens, panels and framing on paid jobs, not a staged lab.
Licensed trades: Electricians, plumbers and carpenters doing the work they're trained for.
Hours, not clips: Whole tasks from setup to clean-up, redos and interruptions included.

Pl. IIThrough the glasses: the wearer's own hands at the box.

What each session becomes

The hands, the steps and the rules.

Every session is turned into what a robot policy learns from: where both hands are in every frame, what they hold, what step they're on and whether the work was done safely.

See a timeline

Engraving of a tradesperson's hand turning a wrench on a pipe fitting, joints marked with dots, beside a robot hand making the same grip.

Hands: Both hands as 21 keypoints per frame, in camera and world, with per-frame validity.
Task timeline: Steps with start and end, tools, hands and order. Doubtful ones are confirmed by the worker.
Practice checks: Whether the work followed the safety rules that apply, each cited to its source.

Pl. IIIThe same grip, by hand and by robot.

Checked in physics

If the hands can't do it in physics, it doesn't ship.

Every session is rebuilt as a MuJoCo scene, and the worker's own hands replay it on the surfaces and objects they touched. No assists, welds or teleports.

How it’s checked

Engraving of a cabinet with a pipe and valve on a workbench, beside a measured wireframe copy of it on which a jointed robot hand turns the valve.

Replayed, not assumed: A session ships only when its hands can do in simulation what the video shows.
Objects get exact: A class placeholder first, then the make and model, each one a new version.
Re-exported, not patched: When what a dataset used changes, the dataset is exported again.

Pl. IVThe cabinet, and its measured twin.

Measured, not guessed

Every number says where it came from.

Sizes, masses, frictions and camera settings each carry a value, a range and a source. You can see which were measured on site and which are still assumptions.

Engraving of an electrician on a stepladder at a breaker panel with a torque screwdriver, its dial gauge drawn large with the needle inside a marked band.

Measured or published: Taken on site, or from the maker's own sheet.
Fitted: Solved from the video and the replay, with the range it was solved to.
Assumed: Marked as assumed, so you know what to test before you rely on it.

Pl. VTorqued to spec, and the gauge to show it.

In development

The first sessions are being recorded.

On job sites with a New Jersey contractor. Every worker signs an agreement covering the use, is paid for each hour and reviews their own sessions first. Ask for a sample when they’re ready.

What a sample holds

Engraving of a workshop at evening: safety glasses, a tool belt and a robot hand at rest on the bench, a door open onto the street.

Pl. VIEnd of the day: the tools down, the session uploaded.