From the street, the only sign I could find that it was Physical Intelligence’s headquarters in San Francisco was the pi symbol on the door, which was a slightly different color than the rest of the building. Once inside, you’ll immediately encounter activity. There’s no reception desk, no logo glowing in the fluorescent lights.
Inside, it’s a huge concrete box, with long blond wooden tables arranged haphazardly, making it a little less solemn. Some are clearly for lunch, dotted with boxes of Girl Scout cookies, jars of Vegemite (I’m Australian here), and small wire baskets stuffed with way too many condiments. The remaining tables tell a completely different story. Many more of them are loaded with monitors, robot spare parts, a tangle of black wires, and fully assembled robotic arms in various states as they try to master the mundane.
During my visit, one arm is folding, or trying to fold, a pair of black pants. It doesn’t work. The other is someone who is trying to turn their shirt inside out with such determination that it suggests that they will succeed not just today, but eventually. Third, you are supposed to quickly peel the zucchini and put the shavings in a separate container. At least the shavings are on track.
“Think of it like ChatGPT, but for robots,” Sergey Levine told me, gesturing to the motorized ballet unfolding across the room. Mr. Levine, an associate professor at the University of California, 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.
He explains that what I’m looking at is the testing phase of a continuous loop. Data is collected at robotic stations here and elsewhere, in warehouses, at home, and wherever teams can set up shop, and the data is used to train general-purpose robot-based models. Once researchers train a new model, they return to these stations for evaluation. The pants folder is someone’s experiment. The same goes for turning your shirt. A zucchini peeler might be testing whether the model can be generalized to a variety of vegetables, learning the basic movements of peeling so that it can successfully handle apples and potatoes it has never encountered before.
The company also operates test kitchens in this building and other locations using off-the-shelf hardware to expose robots to different environments and challenges. There’s a sophisticated espresso machine nearby, and you think it’s for the staff until Levin clarifies, “No, it’s there for the robots to learn.” The frothed latte is all data, not a perk for the dozens of engineers on site who mostly peer into computers and watch mechanized experiments.
The hardware itself is intentionally unassuming. The weapons sell for about $3,500, which includes what Levine described as a “huge markup” from the vendor. If manufactured in-house, material costs would drop to less than $1,000. A few years ago, he says, roboticists would have been shocked by what these robots could do. But that’s the point. Good intelligence makes up for bad hardware.
tech crunch event
boston, massachusetts
|
June 23, 2026
As Levine excused himself, Laci Groom approached me, moving through the space with the purposefulness of someone who had six things going on at once. At 31 years old, Groome still has the freshness of a Silicon Valley wonder boy, earning the title early on by starting his first company in his native Australia at the age of 13 and selling it nine months later (that’s how Vegemite describes it).
When I first spoke to him once while welcoming a few sweatshirt-clad visitors into the building, his immediate response to my request to spend time with him was, “No, I have a meeting.” Now he has about 10 minutes.
Groom found what she was looking for when she started following the academic research coming out of Levine and Chelsea Finn’s lab. Finn, a former doctoral student of Levine’s at Berkeley, now runs his own lab at Stanford University specializing in robotic learning. Their names kept popping up in all sorts of interesting things happening in robotics. When he heard rumors that they might be starting something, he tracked down Karol Hausman, a Google DeepMind researcher and professor at Stanford University who knew Groom was involved. “It was one of those meetings where you walk out and say, ‘This is it.'”
Given his track record, some might wonder why he didn’t become a full-time investor, but Groom never intended to become a full-time investor, he told me. 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 himself. His first robot investment, Standard Bot, took place in 2021 and reintroduced him to the field he loved as a child building LEGO Mindstorms. As he jokes, “I spent a lot more time on vacation as an investor.” But investing was just a means to stay active and meet people, not the end goal. “I was looking for a company for five years before it started post-stripe,” he says. “Good ideas at the right time, with a good team. [that’s] Very rare. It’s all about execution, but you can run with a bad idea like crazy and it’s still a bad idea. ”
The two-year-old company has now raised more than $1 billion, and when asked about its runway, he was quick to clarify that it’s not really on fire. Most of the spending goes to computing. Shortly after, he admitted that with the right conditions and the right partner, he could raise more money. “There’s really no limit to how much money you can put into work,” he says. “There’s always more compute to throw at the problem.”
What makes this arrangement particularly unusual is what Groom hasn’t offered his supporters: a timeline for turning physical intelligence into a money-making endeavor. “I don’t answer commercialization to investors,” he says of backers including Khosla Ventures, Sequoia Capital and Thrive Capital, which valued the company at $5.6 billion. “It’s kind of weird that people would tolerate that.” But they tolerate it, and that’s not always the case. That is why it is mandatory for the company to have sufficient capital now.
So what is the strategy if not commercialization? Quan Vuong, another co-founder from Google DeepMind, explains that it revolves around learning beyond the body and diverse data sources. If someone builds a new hardware platform tomorrow, they don’t need to start collecting data from scratch. You can transfer all the knowledge your model already has. “The marginal cost of introducing autonomy to a new robotic platform is much lower, no matter what the platform is,” he says.
The company is already working with a handful of companies in a variety of industries, including logistics, a grocery store, and the chocolate maker across the street, to test whether its systems are good enough for real-world automation. Vuong argues that in some cases this is already the case. With an “any platform, any task” approach, the scope for success is large enough that you can start checking out tasks that are ready for automation today.
Physical intelligence is not the only thing pursuing this vision. Similar to the LLM model that captivated the world three years ago, the race to build general-purpose robot intelligence is the foundation upon which more specialized applications can be built. Skild AI, a Pittsburgh-based company founded in 2023 that just this month raised $1.4 billion at a $14 billion valuation, is taking a markedly different approach. While Physical Intelligence remains focused on pure research, Skild AI has already commercially deployed its “all-purpose” Skild Brain, saying it generated $30 million in revenue across security, warehousing, and manufacturing in just a few months last year.
Skild publicly attacked its competitors, arguing in a blog post that most “robotics foundation models” are just “transformed” visual language models lacking “true physical common sense” because they rely too much on internet-scale pre-training rather than physically-based simulations or real robotics data.
That’s a pretty sharp philosophical divide. Skild AI is betting that commercial deployment will create a data flywheel that improves models for each real-world use case. Physical Intelligence is betting that by resisting the temptations of short-term commercialization, we can create superior general intelligence. It will take years to resolve which is “more correct.”
Meanwhile, physical intelligence operates with what Gloom describes as unusual clarity. “It’s a very pure company. Researchers have needs, and we collect data to support those needs, and we do it with new hardware and whatever it is. It’s not externally driven.” The company had a five- to 10-year roadmap for what the team thought was possible. By 18 months, he says, they had blown it.
The company has about 80 employees and plans to continue growing, but Groom said he hopes to do so “as slowly as possible.” The hardest part, he says, is the hardware. “Hardware is really hard. Everything we do is much harder than a software company.” Hardware breaks. Arrive late and test will be delayed. Everything gets complicated when you consider safety.
I watched the robots continue practicing as the groom stood up and hurried off to his next mission. The pants are not yet folded. The shirt stubbornly remains on the right side out. Zucchini shavings are being piled up nicely.
There are obvious questions, myself included, about whether anyone would actually want to use a vegetable-peeling robot in the kitchen, about safety, about dogs going crazy when machines get into the house, and about whether all the time and money invested here will solve big enough problems or create new ones. Meanwhile, outsiders have questions about the company’s progress, whether its vision is achievable, and whether it makes sense to bet on general intelligence rather than specific applications.
Even if the groom has doubts, he doesn’t show them. He’s working with people who have been working on this problem for decades and believe the timing is finally right, and that’s all he needs to know.
Additionally, Silicon Valley has supported people like grooms since the early days of the industry, giving them a lot of rope. Even if they don’t have a clear path to commercialization, even if they don’t have a timeline, even if they’re not sure what the market will look like when they get there, they know they’re likely to figure it out. It doesn’t always work out. But when we do, we tend to justify many of the times when we don’t.
Source link
