Live now
Crew cameras
Our tradesmen wear first-person cameras on the hard hat. We capture exactly what they see and how they do the job, on real paid projects.
Lab · Data
Data
Real jobs to train on. Real outcomes to test against.
How we capture the trades
Four ways in.
Coming next
Sensor gloves
Gloves that track every finger capture the feel of the work. How hard to press, how to grip a tool, when the concrete is ready.
Coming next
Remote-controlled machines
Skid steers, mini excavators, telehandlers, and power trowels fitted with remote kits. Every command is recorded with camera and sensor data.
Later
Robot arms on our crews
A crew member guides a robot arm on a lift to drive screws and hold trim. Each movement teaches the arm how our crews work.
Verified outcomes
Physical work has no answer key. Ours does.
AI improved fastest at math and code because the answers can be checked. Construction video shows the motion, not whether it was done right. Every Metal America job ends with a verified outcome, including what it cost, how long it took, and whether it passed inspection.
The data pipeline
One backend for every data type
Every modality lands in the same place, gets the same QA and redaction, and is joined to the same job record before anyone labels it. In robotics terms, head cam video and glove data are human demonstrations, and teleoperation produces robot trajectories with real action labels.
- CaptureEpisodes recorded on paid jobs, with written consent from every worker.
- IngestVideo, glove streams, ROS bag and MCAP logs, and LeRobot or RLDS episodes land in object storage. Originals are never changed.
- QA and redactionAutomatic checks for blur, shake, exposure, and hands in frame. Faces are blurred.
- Job linkingEvery episode is joined to its job record, including job type, size, crew ID, materials, hours, cost, inspection result, and callbacks.
- AnnotationAI writes task and keystep labels, and people review the low-confidence segments. How we label
- CurationFiltered and packaged into versioned datasets, each with a data card.
Anchoring framing to the ground
Position
Drill
Anchor
Check
Tied to every job's outcome
Cost, time, and pass or fail on every clip.
That's the part no one else has.
ActionPosition
Job cost$7,920
Duration6h 40m
InspectionPassed
Sample record
Model testing
Test your model on real jobs.
Bring a model. We run it on real-world edge cases from held-out jobs with known outcomes, and show you where it holds up and where it breaks. Live today for language and vision models. Robot testing on job sites comes next.
What's in the dataset
The trades, labeled.
Trades
Concrete, steel erection, framing, wall and roof panels
View
First person, real paid jobs across the US
Labels
Task step, tools, crew size, duration
Outcome
Job cost, schedule, inspection result, callbacks
Labeling
Labels are generated by AI at scale and verified against real job outcomes. How we label
Source
2,437 orders, 9,425 tradesmen, 48 states
Privacy
Crews come first.
- Faces blurred
- Consent from every worker
- Opt out anytime
Steel erection
Questions about verifiable construction data
What is verifiable training data?
Verifiable training data is data where the correct result can be checked against a real outcome, so a model learns what worked, not just what happened.
Why did AI improve fastest at math and coding?
Because the answers can be checked automatically. A proof is right or wrong and code passes its tests or fails, so models can be trained against a reliable signal.
Why is robot training data for construction hard to get?
Physical work usually has no answer key. Video of a task shows the motion, but not whether it was done right. Most footage has no outcome attached.
Can't anyone collect construction data?
Anyone can film a job or build a simulation. Real-world edge cases only come from real jobs at scale, tied to what actually happened. We capture them on every job we build, and we know the outcome.
What outcome is attached to Metal America data?
Every job ends with a verified outcome from our records, including what it cost, how long it took, whether it passed inspection, and whether there was a callback.
Licensing
Two ways to license
Research license
For academic and non-commercial research. Use the data to train, evaluate, and publish.
- Labeled first-person episodes
- Job outcome labels
- Data card and consent records
Commercial license
For companies training or evaluating production robot models and construction AI.
- Everything in research
- Commercial training and evaluation rights
- New data as it is captured
Request access
Building AI for job sites?
Train on our data, test your model on real jobs, or both. We will reply within one business day with what we have and what is coming next.
- Data cards and consent records available
- Faces blurred
- Every episode tied to a real job outcome
- Data licenses and model testing
Prefer to talk? Book a demo with one of our experts ↗