From the street, the only indication I’ve found Physical Intelligence’s headquarters in San Francisco is a pi symbol that’s a slightly different color than the rest of the door. When I walk in, I’m immediately confronted with activity. There’s no reception desk, no gleaming logo in fluorescent lights.
Inside, the space is a giant concrete box made slightly less austere by a haphazard sprawl of long blonde-wood tables. Some are clearly meant for lunch, dotted with Girl Scout cookie boxes, jars of Vegemite (someone here is Australian), and small wire baskets stuffed with one too many condiments. The rest of the tables tell a different story entirely. Many more of them are laden with monitors, spare robotics parts, tangles of black wire, and fully assembled robotic arms in various states of attempting to master the mundane.
During my visit, one arm is folding a pair of black pants, or trying to. It’s not going well. Another is attempting to turn a shirt inside out with the kind of determination that suggests it will eventually succeed, just not today. A third — this one seems to have found its calling — is quickly peeling a zucchini, after which it is supposed to deposit the shavings into a separate container. The shavings are going well, at least.
“Think of it like ChatGPT, but for robots,” Sergey Levine tells me, gesturing toward the motorized ballet unfolding across the room. Levine, an associate professor at UC Berkeley and one of Physical Intelligence’s co-founders, has the amiable, bespectacled demeanor of someone who has spent considerable time explaining complex concepts to people who don’t immediately grasp them.
Image Credits:Connie Loizos for TechCrunch
What I’m watching, he explains, is the testing phase of a continuous loop: data gets collected on robot stations here and at other locations — warehouses, homes, wherever the team can set up shop — and that data trains general-purpose robotic foundation models. When researchers train a new model, it comes back to stations like these for evaluation. The pants-folder is someone’s experiment. So is the shirt-turner. The zucchini-peeler might be testing whether the model can generalize across different vegetables, learning the fundamental motions of peeling well enough to handle an apple or a potato it’s never encountered.
The company also operates a test kitchen in this building and elsewhere using off-the-shelf hardware to expose the robots to different environments and challenges. There’s a sophisticated espresso machine nearby, and I assume it’s for the staff until Levine clarifies that no, it’s there for the robots to learn. Any foamed lattes are data, not a perk for the dozens of engineers on the scene who are mostly peering into their computers or hovering over their mechanized experiments.
The hardware itself is deliberately unglamorous. These arms sell for about $3,500, and that’s with what Levine describes as “an enormous markup” from the vendor. If they manufactured them in-house, the material cost would drop below $1,000. A few years ago, he says, a roboticist would have been shocked these things could do anything at all. But that’s the point — good intelligence compensates for bad hardware.
Techcrunch event TechCrunch Founder Summit 2026: Tickets Live On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately
... continue reading