From the street, the only sign of Physical Intelligence’s headquarters in San Francisco is the pie symbol, a slightly different color than the rest of the door. As soon as I walk in I am immediately faced with activity. There is no reception desk, no logo flashing under fluorescent lights.
The interior space is a huge concrete box, slightly less austere, with long blond wooden tables spread out haphazardly. Some are clearly filled with boxes of Girl Scout cookies for lunch, jars of Vegemite (someone here is Australian), and little wire baskets filled with way too many condiments. The rest of the table tells a completely different story. More of them are equipped with monitors, spare robotic parts, a tangle of black wires, and fully assembled robotic arms in various states to master everyday tasks.
During my visit, one arm is folded or folded into a pair of black trousers. It’s not working. Another is to try to turn the shirt inside out with determination that it will eventually succeed. But not today. The third (this one seems to have found its calling) is to quickly peel the pumpkin and place the crumbs in a separate container. At least the crumbs are going well.
“Think of it like ChatGPT, but for robots.” Sergey Levine tells me, pointing to the electric ballet taking place across the room. Levine, an associate professor at UC Berkeley and one of the co-founders of Physical Intelligence, has the affable, bespectacled demeanor of someone who has spent considerable 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 from robotic stations here and in other locations (warehouses, homes, wherever the team might set up shop), and that data trains a general-purpose robot-based model. As researchers train new models, they return to stations like these for evaluation. The pants folder is someone’s experiment. The same goes for Shirt Turner. A pumpkin peeler can test whether a model can generalize across a variety of vegetables and learn the basic behavior of peeling enough to handle an apple or potato it’s never seen before.
The company also operates test kitchens in this building and elsewhere using off-the-shelf hardware to expose the robots to a variety of environments and challenges. There’s a sophisticated espresso machine nearby, which I assume is for staff use until Levine clarifies that no, the robot is there to learn. A frothy latte is mostly just data, not a perk for the dozens of engineers in the field peering at computers or hovering over mechanized experiments.
The hardware itself is intentionally unattractive. The weapon sells for about $3,500, which Levine describes as a “huge markup” from the supplier. If made in-house, material costs would drop to less than $1,000. Just a few years ago, he says, roboticists would have been shocked that anything like this could be done. But that’s the point. Good intelligence compensates for bad hardware.
Tech Crunch Event
Boston, Massachusetts
|
June 23, 2026
As Levine excused himself, Lachy Groom approached me, moving through the space with the purposefulness of someone who had six things happening at once. At 31, Groom still has the fresh air of a Silicon Valley boy wonder, a designation he earned early on after starting his first company in his native Australia at age 13 and selling it nine months later (which explains Vegemite).
When I first approached him, as he welcomed the little visitors in sweatshirts to the building, he immediately responded to my request to spend time with him: “Absolutely not. I have a meeting.” I think he has about 10 minutes now.
Groom found what he was looking for when he began following academic research coming out of the lab of Levine and Chelsea Finn, Levine’s former Berkeley doctoral student who now runs her own lab at Stanford focused on robotic learning. Their names keep popping up on all the exciting things happening in robotics. When he heard a rumor that they might be starting something, he tracked down Karol Hausman, a Google DeepMind researcher who taught at Stanford and whose involvement Groom had learned of. “It was one of those meetings where you just go out and think, ‘This is it.’”
The groom said he never intended to become a full-time investor. Some people may wonder why they don’t disclose their performance. After leaving Stripe, where he was an early employee, he spent about five years as an angel investor, making early investments in companies like Figma, Notion, Ramp, and Lattice, looking for the right company to start or join. His first robotics investment, Standard Bots, came in 2021 and reintroduced him to his childhood love of building LEGO Mindstorms. As he joked, he was “much more on vacation as an investor.” But investing is just a way to stay active and meet people, not the end game. “I was looking for five years for a company to start after Stripe,” he says. “It’s very rare that a good idea comes at a good time with a good team. It’s all just execution, but a bad idea can be executed like hell and it’s still a bad idea.”
The two-year-old company has now raised more than $1 billion, and when I ask him about the runway, he’s quick to clarify that it doesn’t actually consume that much. The majority of spending is on computing. After a while, he acknowledged that he could raise more money under the right conditions and with the right partners. “There’s really no limit to how much we can invest in our work,” he says. “There is always more computing available to solve problems.”
What makes this arrangement particularly unusual is that Groom does not provide backers with a timeline for turning their physical intelligence into a money-making endeavor. “I don’t give investors answers about commercialization,” he says of backers including Khosla Ventures, Sequoia Capital and Thrive Capital, which value the company at $5.6 billion. “It’s so strange that people put up with that.” But you may condone what they do and that may not always be the case. That’s why it makes sense for the company to be well capitalized now.
So what is the strategy if not commercialization? Quan Vuong, another co-founder from Google DeepMind, explains that it revolves around cross-implementation learning and diverse data sources. If someone were to build a new hardware platform tomorrow, there would be no need to start collecting data from scratch. Any knowledge the model already has can be transferred. “The marginal cost of giving autonomy to a new robotics platform, whatever the platform, is much lower,” he says.
The company is already working with a handful of companies in a variety of industries, including logistics, grocery and the chocolate manufacturer across the street, to test whether their systems are good enough for real-world automation. Vuong argues that in some cases it already is. With an “any platform, any task” approach, the surface area for success is large enough to start identifying tasks that are ready to be automated today.
Physical intelligence is not the only thing driving this vision. The race to build general-purpose robotic intelligence, a foundation on which more specialized applications can be built, like the LLM model that took the world by storm three years ago, is heating up. Pittsburgh-based Skild AI, founded in 2023 and raising $1.4 billion this month at a $14 billion valuation, is taking a markedly different approach. Although Physical Intelligence still focuses on pure research, Skild AI has already deployed its “omnibody” Skild Brain commercially and said last year that it had generated $30 million in revenue in just a few months across security, warehouse and manufacturing.
Skild has even publicly criticized competitors on his blog, claiming that most “robotics-based models” are just “disguised” vision language models that lack “true physical common sense” because they rely too heavily on physics-based simulations and Internet-scale pre-training rather than real-world robotics data.
That’s a pretty sharp philosophical divide. Skild AI is confident that commercial deployment will generate a data flywheel that improves models for each real-world use case. I am confident that Physical Intelligence will be able to produce superior general intelligence if it resists short-term commercialization. It will take years to determine who is “righter.”
In the meantime, body intelligence comes into play very clearly, as Groom explains. “It’s a really pure company. Researchers have a need. We collect data, new hardware, whatever, to support that need. It’s not externally driven.” The company had a 5-10 year roadmap for what the team thought was possible. By 18 months, they had completely passed it, he said.
The company has about 80 employees and plans to grow, but the groom said he hopes to do so “as slowly as possible.” The hardest thing, he says, is the hardware. “Hardware is really hard. Everything we do is much harder than a software company.” The hardware is damaged. It arrives slowly, causing delays in testing. Safety considerations complicate everything.
I was watching the robots continue practicing while my husband ran to his next appointment. The pants are not fully folded yet. The shirt stubbornly remains right side out. Pumpkin peels are piled up nicely.
There are obvious questions, including my own, about whether anyone would actually want a robot peeling vegetables in their kitchen, about safety, whether dogs go crazy seeing machines breaking into their homes, and whether all the time and money invested in it solves a big enough problem or creates new ones. Meanwhile, outsiders question the company’s progress, whether it can achieve its vision, and whether it makes sense to bet on general intelligence rather than specific applications.
If the groom has any doubts, don’t show them. He’s working with people who have been working on this problem for decades and who believe the time has finally come. This is all he needs to know.
Moreover, Silicon Valley supported people like Groom early in the industry and gave them a lot of rope. Even without a clear path to commercialization, without a timeline, without certainty about what the market will look like, they know they have a good chance of figuring it out. It doesn’t always work. But that tends to justify a lot of things that aren’t the case.
