Physical Intelligence is working on one of Silicon Valley’s most ambitious robotics ideas: software that can give robots broader, more transferable intelligence. Inside its San Francisco headquarters, robotic arms practice ordinary tasks such as folding pants, turning a shirt inside out, peeling a zucchini, and learning around an espresso machine.
The company’s bet is that better intelligence can make modest hardware much more capable. That puts Physical Intelligence in the middle of a growing race to build robot brains that can move beyond narrow, single-purpose automation.
A lab built around everyday robot practice
Physical Intelligence’s headquarters is not presented as a polished showroom. The space is described as a large concrete room filled with long blonde-wood tables, food items, robotics parts, wires, monitors, and robotic arms in different stages of testing.
The scenes are deliberately ordinary. One robot arm is trying to fold black pants. Another is working on turning a shirt inside out. A third is peeling a zucchini and is meant to place the shavings into a separate container.
Sergey Levine, an associate professor at UC Berkeley and one of the company’s co-founders, describes the idea simply: “Think of it like ChatGPT, but for robots.” The comparison points to the company’s goal of creating general-purpose robotic foundation models rather than training a robot for one isolated job.
The work runs in a loop. Robots collect data at stations in the office and at other locations, including warehouses and homes where the team can set up. That data is then used to train models, which return to test stations for evaluation.
Why cheap hardware matters
The robotic arms in use are intentionally not expensive industrial showpieces. Levine says the arms sell for about $3,500, including what he calls “an enormous markup” from the vendor. If Physical Intelligence made them in-house, he says the material cost would drop below $1,000.
That choice matters because it frames the company’s core belief. The hardware does not have to be perfect if the intelligence controlling it becomes strong enough. Levine says a roboticist a few years ago would have been surprised that such machines could do anything useful at all.
The company also uses a test kitchen in the building and elsewhere. The espresso machine nearby is not just for employees. It is there so robots can learn from the task environment, meaning even foamed lattes can become training data.
This approach puts emphasis on variety. The zucchini-peeling task, for example, may help test whether a model can generalize the motions of peeling across different foods, including an apple or a potato it has not encountered before.
Lachy Groom’s long search for the right company
Lachy Groom is central to the company’s story. At 31, he has already built a reputation in Silicon Valley. He sold his first company nine months after starting it at age 13 in Australia, and later became an early employee at Stripe.
After Stripe, Groom spent roughly five years as an angel investor. His early bets included Figma, Notion, Ramp, and Lattice. He says investing was not the final plan, but a way to stay active and meet people while looking for the right company to start or join.
His first robotics investment was Standard Bots in 2021. That brought him back to a field he had enjoyed as a child building Lego Mindstorms. Groom says he was “on vacation much more as an investor,” but he was still looking for a company with the right timing, idea, and team.
He began following academic work from 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 robotic learning. Their names kept appearing in robotics work that interested him. When he heard they might be starting something, he tracked down Karol Hausman, a Google DeepMind researcher who also taught at Stanford and was involved. Groom later said, “It was just one of those meetings where you walk out and it’s like, This is it.”
Big funding without a fixed commercial clock
Physical Intelligence is two years old and has raised over $1 billion. Groom says most of its spending goes toward compute, and he is open to raising more under the right terms and with the right partners. As he puts it, “There’s no limit to how much money we can really put to work,” because “There’s always more compute you can throw at the problem.”
The company’s backers include Khosla Ventures, Sequoia Capital, and Thrive Capital, which have valued the company at $5.6 billion. What they do not get, according to Groom, is a clear commercialization timeline. “I don’t give investors answers on commercialization,” he says. “That’s sort of a weird thing, that people tolerate that.”
The strategy instead centers on research and transferability. Quan Vuong, another co-founder who came from Google DeepMind, says the company is focused on cross-embodiment learning and diverse data sources. The aim is that a new hardware platform would not have to begin data collection from zero, because it could use knowledge the model already has.
Vuong says, “The marginal cost of onboarding autonomy to a new robot platform, whatever that platform might be, it’s just a lot lower.” Physical Intelligence describes its approach as “any platform, any task.”
The robotics foundation model race
Physical Intelligence is already working with a small number of companies in different verticals, including logistics, grocery, and a chocolate maker across the street. Vuong says that in some cases the systems are already good enough for real-world automation.
But the company is not alone. Skild AI, based in Pittsburgh and founded in 2023, is taking a different route. It raised $1.4 billion at a $14 billion valuation and has already deployed its “omni-bodied” Skild Brain commercially. Skild AI says it generated $30 million in revenue in just a few months last year across security, warehouses, and manufacturing.
Skild AI has also argued on its blog that many “robotics foundation models” are only vision-language models “in disguise” and lack “true physical common sense” because they depend too much on internet-scale pretraining rather than physics-based simulation and real robotics data.
The divide is clear. Skild AI is using commercial deployment as a way to create more real-world data. Physical Intelligence is holding back from near-term commercialization because it believes that focus may produce stronger general intelligence over time.
Groom says Physical Intelligence once had a 5- to 10-year roadmap, but by month 18 the team had already moved through it. The company has about 80 employees and plans to grow, though he says hopefully “as slowly as possible.” The hardest part, he says, is hardware: “Hardware is just really hard. Everything we do is so much harder than a software company.”
The open questions remain large. It is not yet clear how much demand there will be for robots in homes, kitchens, warehouses, or other settings, or how safety and practical deployment will shape the market. For now, Physical Intelligence is betting that the right robot brains must come before the clearest business model.