Built Robotics and Penn's Safe Autonomous Systems Lab want to train machines to handle the messy reality of human crews on a job site, not just controlled test courses.
On a utility-scale solar site somewhere in the American Southwest, a robotic pile driver called the RPD 35 is sinking steel supports into hardpan dirt while human crews work within the same footprint: checking alignments, routing conduit, keeping clear of the machine's swing radius. Built Robotics says it has logged more than 50,000 hours of exactly this kind of work, across 40 or more active sites, with a company tally of more than 3 gigawatts of solar installed, all numbers the source describes as Built's own claims rather than independently verified The Robot Report.
The next problem is harder. Built's autonomous equipment can follow a programmed piling pattern, but it does not yet have a working model of the people, machines, and shifting terrain around it. That gap is what a newly announced partnership with the Safe Autonomous Systems Lab (xLAB) at the University of Pennsylvania's engineering school is meant to close. The plan, as the companies describe it, is to put a small mobile robot fitted with a sensor suite on active job sites. Researchers at xLAB will then turn the resulting data into what the partnership calls a "world foundation model": a model of how people, machines, and terrain interact on a real site, not a virtual world or a simulation.
xLAB is led by Rahul Mangharam, a professor of electrical and systems engineering at Penn. In announcing the partnership, he framed the bet in unusually direct terms. "Physical AI," the broad label for AI that acts on equipment in the real world rather than on screens, "has shown impressive performance in controlled environments," Mangharam said, "but bridging the gap between validation in controlled environments and robust performance under operational conditions remains a central challenge" The Robot Report. That sentence is the partnership's own critical frame, an admission that the field's biggest claim has not yet held up outside the lab.
Built's founder and CEO, Noah Ready-Campbell, is a Penn engineering alum, which gives the deal a direct institutional link. The company has been operating in construction since 2016 and entered the utility-scale solar market in 2023 with the RPD 35, an autonomous pile driver built for the high-volume, repetitive work that solar farms require. "Physical AI cannot advance without data from real, operational environments," Mangharam said, and Built is one of the few companies that has it, in the form of years of piling runs on real dirt The Robot Report.
"Physical AI" and "world foundation model" are the partnership's own framing, not an established industry category, and the source does not specify funding, intellectual property ownership, deliverable timeline, deployment scope, or commercialization path for the foundation model. What the announcement does name is the central technical problem, and the central technical problem is the one the industry has been working on for years: a model trained on real, messy job sites can still fail in a way a model trained in a controlled environment never sees, and the failure mode is the one that matters most when a human crew is working in the same frame.
What to watch next is whether the data-collection robot actually shows up on Built's active sites, what kind of incidents or near-misses its scans capture, and whether xLAB's safety research produces a model that can be tested in the field rather than only in a paper. A clean replication of operational conditions in a lab is not the same thing as safe performance on a site where a human crew is working three meters from a swinging arm. The partnership is, in effect, a bet that the two can be made to converge, and the next twelve months on Built's solar sites will be the first real read on whether the data, the lab, and the equipment can be made to fit.