From the street, the only indication I found of Physical Intelligence headquarters in San Francisco was a pi symbol in a slightly different color than the rest of the door. When I enter, I am immediately met with activity. There is no reception, no logo glowing in the fluorescent lights.
Inside, the space is a giant concrete box made slightly less austere by a haphazard spread of long, blond wooden tables. Some are clearly intended for lunch, dotted with Girl Scout cookie tins, jars of Vegemite (someone here is Australian) and small wire baskets filled with way too many condiments. The rest of the paintings tell a completely different story. Many more of them are loaded with monitors, spare robotic parts, tangles of black wires, and fully assembled robotic arms in various states in an attempt to master the mundane.
During my visit, an arm is folding black pants, or trying to do so. It’s not going well. Another tries to turn a shirt inside out with the kind of determination that suggests she will eventually succeed, but not today. A third – this one seems to have found his vocation – quickly peels a zucchini, after which he is supposed to place the shavings in a separate container. The shavings are fine, at least.
“Think of it like ChatGPT, but for robots,” Sergey Levine tells me, pointing to the motorized ballet unfolding across the room. Levine, an associate professor at UC Berkeley and one of the co-founders of Physical Intelligence, has the amiable, bespectacled look of someone who has spent a lot of time explaining complex concepts to people who don’t immediately understand them.
What I’m seeing, he explains, is the testing phase of a continuous loop: Data is collected at robotic stations here and in other locations – warehouses, homes, wherever the team can set up shop – and that data trains basic general-purpose robotic models. When researchers train a new model, it returns to stations like these for evaluation. The pants folder is someone’s experience. The shirt turner too. The zucchini peeler could test whether the model can generalize to different vegetables, learning the fundamental peeling motions well enough to handle an apple or potato it has never encountered.
THE business also operates a test kitchen in this building and elsewhere using commercially available hardware to expose robots to different environments and challenges. There’s a fancy espresso machine nearby, and I assume it’s for the staff until Levine points out that no, it’s there for the robots to learn. All the foamed slats are data, not a benefit to the dozens of engineers on site, many of whom peer into their computers or hover over their mechanized experiments.
The material itself is deliberately unglamorous. These guns sell for about $3,500, which Levine describes as “a huge markup” from the seller. If they made them in-house, the cost of materials would drop below $1,000. A few years ago, he says, a roboticist would have been shocked to see that these machines could do anything. But that’s the point: good intelligence makes up for bad hardware.
Techcrunch event
Boston, Massachusetts
|
June 23, 2026
As Levine excuses himself, I’m approached by Lachy Groom, who moves through the space with the determination of someone doing half a dozen things at once. At 31, Groom still has the fresh face of the Silicon Valley wunderkind, a title he earned early, having sold his first company nine months after starting it at age 13 in his native Australia (that explains the Vegemite).
When I first approached him earlier, as he welcomed a small group of sweatshirt-clad visitors into the building, his response to my request for time with him was immediate: “Absolutely not, I have appointments.” He now has maybe 10 minutes.
Groom found what he was looking for when he began following the academic work of the labs of Levine and Chelsea Finn, a former doctoral student of Levine’s at Berkeley who now runs her own lab at Stanford focused on robotic learning. Their names kept popping up in everything interesting happening in robotics. When he heard rumors that they might start something, he tracked down Karol Hausman, a Google DeepMind researcher who also taught at Stanford and whom Groom had learned was involved. “It was just one of those meetings where you walk out and you’re like, This is it.”
Groom never intended to become a full-time investor, he tells me, although some might wonder why he didn’t do so given his background. After leaving Stripe, where he was an early employee, he spent about five years as an angel investor, making early bets on companies like Figma, Notion, Ramp and Lattice while looking for the right company to start or join. His first investment in robotics, Standard Bots, took place in 2021 and reintroduced him to a field he loved as a child, building Lego Mindstorms. As he jokes, he was “much more on vacation as an investor.” But investing was just a way to stay active and meet people, not an end goal. “I was waiting five years for the company to start after Stripe,” he says. “Good ideas at the right time with a good team — [that’s] extremely rare. It’s all execution, but you can execute like hell on a bad idea, and it’s still a bad idea.
The company, created two years ago, has now raised more than a billion dollarsand when I ask about his track, he is quick to point out that it doesn’t really burn much. Most of its spending is on computing. A moment later, he recognizes that under the right conditions, with the right partners, he would harvest more. “There’s no limit to how much money we can actually spend,” he says. “There are always more calculations you can do to solve the problem.”
What makes this arrangement particularly unusual is what Groom doesn’t give his backers: a timetable for turning physical intelligence into a lucrative business. “I don’t give answers to investors about commercialization,” he says of backers such as Khosla Ventures, Sequoia Capital and Thrive Capital, among others, who have valued the company at $5.6 billion. “It’s a little weird that people tolerate this.” But they tolerate it, and that may not always be the case, which is why it’s incumbent on the company to be well capitalized now.
So what is the strategy, if not marketing? Quan Vuong, another co-founder from Google DeepMind, explains that it revolves around cross-incarnation learning and various data sources. If someone builds a new hardware platform tomorrow, they won’t need to start data collection from scratch: they can transfer all the knowledge the model already has. “The marginal cost of integrating autonomy onto a new robotics platform, whatever that platform may be, is simply much lower,” he says.
The company is already working with a small number of companies in different industries (logistics, grocery, chocolate shop across the street) to test whether their systems are good enough for real automation. Vuong says in some cases they already are. With their “any platform, any task” approach, the surface area for success is large enough to start ticking off tasks that are ready for automation today.
Physical intelligence is not the only one pursuing this vision. The race to build general-purpose robotic intelligence – the foundation on which more specialized applications can be built, like the LLM models that captivated the world three years ago – is intensifying. Pittsburgh-based Skild AI, founded in 2023, this month raised $1.4 billion in a Valuation of $14 billion and takes a significantly different approach. While physical intelligence remains focused on pure research, Skild AI has already commercially deployed its “omnibody” Skild Brain, claiming to have generated $30 million in revenue in just a few months last year in security, warehouse and manufacturing.
Skild even took public photos of competitors, argue on your blog that most “basic robotics models” are just “disguised” vision language models that lack “real physics common sense” because they rely too much on internet-scale pre-training rather than physics-based simulation and real robotics data.
This is a fairly stark philosophical gap. Skild AI is betting that commercial deployment creates a flywheel of data that improves the model with each real-world use case. Physical intelligence is betting that resisting the lure of short-term commercialization will enable it to produce superior general intelligence. It will take years to determine who is “most right.”
Meanwhile, physical intelligence operates with what Groom describes as unusual clarity. “It’s such a pure business. A researcher has a need, we collect data to meet that need – or new hardware or whatever – and then we do it. It’s not externally driven.” The company had a 5-10 year roadmap of what the team thought was possible. By the 18th month, they were past it, he said.
The company has about 80 employees and plans to grow, although Groom says she hopes “as slowly as possible.” The biggest challenge, he says, is the hardware. “Hardware is really, really hard. Everything we do is way harder than a software company.” Hardware Breaks. It arrives slowly, delaying the tests. Security considerations complicate everything.
As Groom rushes to rush to his next engagement, I find myself watching the robots continue their training. The pants aren’t quite folded yet. The shirt stubbornly stays right side out. The zucchini shavings accumulate well.
There are obvious questions, including my own, about whether anyone really wants a robot to peel vegetables in their kitchen, about safety, about dogs going crazy over mechanical intruders in their homes, about whether all the time and money invested here is solving big enough problems or creating new ones. Meanwhile, outsiders question the company’s progress, whether its vision is achievable, and whether it makes sense to focus on general intelligence rather than specific applications.
If Groom has doubts, he doesn’t show them. He’s working with people who have been working on this problem for decades and think the time has finally come, that’s all he needs to know.
Additionally, Silicon Valley has been supporting people like Groom and giving them a lot of support since the beginning of the industry, knowing that there’s a good chance that even without a clear path to commercialization, even without a timeline, even without certainty about what the market will look like once they get there, they’ll figure it out. This doesn’t always work. But when it does, it tends to justify most of the time what was not the case.
