Self-driving won on the fleet, not the robot. Construction AI will too.
By Tyler HurdUpdated
Definition. Physical AI is artificial intelligence that perceives and acts in the physical world through a body, such as a car, a robot arm, or a humanoid.
In 2026, one kind of physical AI actually works at scale. Cars drive themselves on public roads. Everything else, humanoids, warehouse arms, construction robots, is still mostly demos.
Self-driving didn't win because of one breakthrough. It won because seven things were built, in order, and each one made the next possible. We think construction follows the same path. Metal America has already built the first four.
1234567
4 of 7 in place
Figure 1. The seven pillars as a dependency graph. Each pillar needs the ones before it. Self-driving has all seven. Metal America has the first four, and the next one is already starting.
What did Tesla actually win with?
Tesla's edge was never one model. It was the data engine. Every time a driver took over, the car flagged an edge case, and the fleet sent back more clips like it. New models ran silently in shadow mode and were judged against what the human driver actually did. The fleet is the moat. Models get replaced every year. Real-world edge cases don't.
Why does the order of the pillars matter?
Cars had to be sold, then driven on real routes, stopping at real stop signs and red lights, before any of that video meant anything. The footage was only valuable because it was attached to real driving with real outcomes.
Construction is the same. Jobs have to be sold and physically completed before a camera on a crew teaches a machine anything. That's the hard, unglamorous part most robotics startups skip. We did it first.
The seven pillars
Self-driving vs Metal America
Progress bars are our own estimates.
01
Fleet
Have
Self-driving
Millions of cars sold to real customers, driving real roads every day.
Metal America
Real customers paying for real buildings and slabs, in 48 states.
2,437 orders and 9,425 tradesmen across 48 states on the Index. About $5M a month in sales.
02
Real-world operation
Have
Self-driving
Cars had to drive true routes, stop at stop signs and red lights before any footage meant anything.
9,425 tradesmen in our network across 48 states, mapped on the Index.
03
Ground-truth data engine
Have
Self-driving
Every driver takeover flags an edge case, and the fleet sends back more like it.
Metal America
Every failed inspection, callback, and delay flags an edge case, tied to the footage and the outcome.
Five AI models label our job site photos, and their majority vote is 98% accurate. When all five disagreed with our label, it was our label that needed a second look. AI data labeling →
04
Planning
Have
Self-driving
Decide what happens next, in order, safely.
Metal America
Know every step of an install, what comes next, and where it fails.
40 install steps mapped across slab and building, from real jobs, with order, tools, time, and failure points. The Install Sequence Graph →
05
Sensors at scale
Started
Self-driving
Eight cameras on every car in the fleet.
Metal America
First-person cameras on every crew, then gloves and machine telemetry.
Crew cameras are live on our first crews. The goal is 100,000 crews.
06
Model testing
Started
Self-driving
New models run in shadow mode and are judged against what the driver actually did.
Metal America
Any model runs on our jobs and is judged against what the crew did and how the job turned out.
Six language models tested on 387 real construction questions, from 19% correct alone to 91% with our data. Construction AI Eval → Five frontier models named the construction stage in our job photos 88 to 100% of the time. Can AI judge construction quality? →
07
Deployment
Next
Self-driving
The car drives itself.
Metal America
Models and robots that pass go to work on our job sites. We don't build them, we prove them.
Remote kits for skid steers, telehandlers, and trowels come first. See Data.
Where does Metal America sit?
Tesla had to build all seven pillars itself. Construction doesn't need one company to. Model makers and robot builders own the middle. We own both ends, the fleet that finds edge cases and the jobs that judge what works.
What was Tesla's training bottleneck?
On Tesla's October 2023 earnings call, Elon Musk said training was "the fundamental limiting factor" for progress in full self-driving. Three years later, video models label and reason over footage at a fraction of the cost, so compute is now cheap. Real-world edge cases aren't.
What are the last three pillars?
Sensors at scale. Cameras on every crew, then gloves and machine telemetry. Our first crews are already wearing them, and capture runs on every job as part of how we build.
Model testing. Any model, from any lab, run against real jobs with known outcomes. We already do this for language and vision models. See our research.
Deployment. Models and robots that pass go to work on our job sites, starting with remote-controlled machines. Tradesmen become operators.
None of these three is a research problem anymore. The bottleneck is data, and real job site data only comes from real jobs.
Why is construction next for physical AI?
Most physical AI teams start with the robot and look for job sites later. We started with the job sites. Data first, robots second, and every model tested on real work.
Physical AI is AI that perceives and acts in the physical world through a body, such as a self-driving car, a robot arm, or a humanoid robot.
Why did self-driving succeed before other physical AI?
Self-driving built seven things in order: a fleet of customers, real-world operation, a ground-truth data engine that finds edge cases, planning, sensors at scale, model testing in shadow mode, and deployment. Each made the next one possible.
What does construction robotics need?
The same seven pillars. Real jobs and crews have to exist before camera footage of the work becomes useful training data.
Which pillars does Metal America have?
Four of seven: a national customer base, crews completing real jobs, outcome data on every job, and planning. Model testing has started. Sensors at scale and deployment are next.
Tyler leads Metal America's AI agents, the LLM systems that talk with thousands of construction buyers. He works daily on prompting, evaluation, and deploying language models in production.
