Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
From the street, the only sign that I’ve found the headquarters of Physical Intelligence in San Francisco is the pi symbol that’s slightly different from the other door. When I enter, I am immediately confronted with events. No reception desk, no logo glowing in fluorescent lights.
Inside is a large concrete box made slightly less rigid by the random stack of long blonde wooden tables. Some are made for lunch, with Girl Scout cookie boxes, Vegemite jars (someone here is Australian), and little wire baskets filled with lots of goodies. The rest of the tables tell a different story. Many of them are burdened with detectors, robotic transformers, black wire, and well-connected robotic devices in various states trying to figure out the commons.
During my trip, one arm is folding the black pants, or trying to. It’s not going well. Someone is trying to turn the shirt inside out with a determination that shows they will eventually win, just not today. The third – this one seems to have found its calling – is chasing the zucchini, after which it should put the shavings in a separate container. The shave is going well, at least.
“Think of this as ChatGPT, but with robots,” Sergey Levine tells me, looking at the ballet of fire going on in the room. Levine, an associate professor at UC Berkeley and one of the founders of Physical Intelligence, has the fun, bright personality of someone who has spent a lot of time explaining complex concepts to people who don’t immediately understand them.
What I’m looking for, he explains, is the experimental part of a continuous cycle: data is collected on the robotics site here and in other places – warehouses, buildings, wherever the group can set up shop – and this data trains models of the basis of robotic objectives. When researchers train a new model, it goes back to places like this for further analysis. Pants-folder is someone’s experiment. So is the shirt turner. The zucchini-peeler can test whether the model can combine different vegetables, learning the basics of peeling well to handle an apple or potato that it has never encountered before.
The company also operates a test kitchen in the building and elsewhere using off-the-shelf equipment to expose robots to different environments and challenges. There’s an advanced espresso machine nearby, and I think it’s for the workers until Levine explains that no, it’s there for the robots to learn. Every frothy latte is data, not the benefit of the many engineers on site who are usually staring at their computers or focused on their mechanical experiments.
The hardware itself is deliberately unglamorous. These arms sell for around $3,500, and that’s what Levine describes as a “great number” from the dealer. If they make it in-house, the cost could drop below $1,000. A few years ago, he says, a roboticist would have been amazed that these things could do anything. But that’s the point – good intelligence pays for bad equipment.
Techcrunch event
Boston, MA
| |
June 23, 2026
As Levine excuses himself, I approach Lachy Groom, moving through the air with the intent of a man who has half of things going on at once. At 31 years old, Mkwati still has the fresh look of a Silicon Valley boy, a name he earned early, after selling his first company nine months after starting it at age 13 in Australia (this explains Vegemite).
When I arrived at his place earlier, when he welcomed a small group of sweat-clad guests into the house, his immediate response to my request to be with him was: “No, I have meetings.” Now he has ten minutes, maybe.
The groom found what he was looking for when he began pursuing a career in the labs of Levine and Chelsea Finn, a former Berkeley PhD student of Levine’s who now runs her own lab at Stanford focused on robotics. Their names still appear in all the exciting developments in robotics. After hearing rumors that he might be starting something, he sought out Karol Hausman, a Google DeepMind researcher who also taught at Stanford and who Groom studied with. “It was one of those meetings where you walk out and it’s like, This is it.”
Groom didn’t want to be a full-time businessman, he tells me, though some might wonder why he didn’t give his background away. After leaving Stripe, where he was an early employee, he spent nearly 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 robotic investment, Standard Bots, arrived in 2021 and brought him back to his childhood favorite of building Lego Mindstorms. As a joke, he was “on vacation as much as a businessman.” But investing was only a way to encourage and meet people, not an end in itself. “I wanted five years to start the company after Stripe,” he says. “A good idea at a good time and a good team – (that’s) very rare. It’s all execution, but you can do as hell on a bad idea, and it’s still a bad idea.”
The two-year-old company has now upgraded more than $1 billionand when I ask him about his runway, he’s quick to clarify that it’s not too hot. Most of his money goes to computing. A moment later, he admits that under the right words, with the right partners, he raises more. “There is no limit to the amount of money we can spend,” he said. “There are always more computers that you can handle.”
What makes this plan so unusual is what Groom doesn’t give his sponsors: time to turn Physical Intelligence into a money-making project. “I don’t offer business solutions for businesses,” he says of backers that include Khosla Ventures, Sequoia Capital and Thrive Capital among others that have valued the company at $5.6 billion. “It’s a wonderful thing, that people tolerate it.” But they are tolerant, and that may not always be the case, which is why the company should have more money now.
So what is the strategy, if not marketing? Quan Vuong, another co-founder who came from Google DeepMind, explains that it is related to learning from different sources and different data sources. If someone develops a new hardware platform tomorrow, they won’t need to start data collection from scratch – they can transfer all the information the brand already has. “The minimum cost of independence for a new robot platform, no matter what the platform is, is very low,” he says.
The company is already working with a few companies in different sectors – goods, groceries, chocolate makers across the street – to test if their systems are good enough to create real-world machines. Vuong says that sometimes they are already there. With their “every platform, every job” approach, the best place for success is to start looking at automated jobs today.
Physical Intelligence is not alone in chasing this vision. The race to build robotics intelligence – the foundation upon which specialized programs can be built, such as the LLM models that captivated the world three years ago – is burning. Pittsburgh-based Skild AI, which was founded in 2023, this month just raised $1.4 billion The cost of $ 14 billion and they are taking a very different approach. Although Physical Intelligence is still focused on pure research, Skild AI has already deployed its Skild Brain “with a full body”, saying that it made an investment of $ 30 million in a few months last year across security, storage, and manufacturing.
Skild fired in the crowd at his competitors, debate on his blog that “models of the basis of robotics” are only examples of the language of vision “hidden” that do not have “real intelligence” because they rely more on online training than on simulations using physics and real data of robotics.
It is a beautiful division of philosophy. Skild AI is betting that commercial shipping creates a flywheel of data that makes the brand more resilient to any real-world problem. Physical Intelligence is betting that resisting the pull of a close trade will make it more intelligent. Who is ‘most right’ will take years to resolve.
Meanwhile, Physical Intelligence deals with what Groom describes clearly. “It’s a clean company. A researcher has a need, we go and collect data to support that need – or new hardware or whatever it is – and then we do it. It’s not controlled by the outside.” The company had a 5- to 10-year road map that the team thought would be finished. By the 18th month, they had shot, he says.
The company has about 80 employees and plans to grow, though Groom says hopefully “as slowly as possible.” The hardest part, he says, is the hardware. “Hardware is more complex. Everything we do is more complex than a software company.” Hardware explosion. It comes slowly, delaying the exam. Security considerations interfere with everything.
While Groom started to rush off to another volunteer, I was left watching the robots continue their work. The pants are still unrolled. The shirt remains stuck right-side-out. The zucchini legs are stacking well.
There are obvious questions, including mine, if everyone really wants a robot in their kitchen peeling vegetables, about security, about dogs going crazy over robots in their homes, if it’s always money aside, they solve big enough problems or create new ones. Meanwhile, outsiders question the company’s progress, whether its vision is feasible, and whether it is betting on common sense rather than practical application.
If the Bridegroom has any doubt, he does not show it. They are working with people who have been doing this for years and believe that the timing is right, which is what they need to know.
After all, Silicon Valley has been supporting people like the Bridegroom and giving them many ropes from the beginning of the business, knowing that there is a chance that even without a well-known way to go to business, even without time, even without being sure what the market will look like when they get there, they will see. It’s not always good. But when it does, it tends to justify more often than not.
