From the street, the only hint of Physical Intelligence’s headquarters in San Francisco is a pi symbol that’s a slightly different color than the rest of the door. When I step inside, I’m immediately struck by the activity. There’s no reception desk or flashy logo lighting up the space.
The interior resembles a massive concrete box, made a bit more inviting by an assortment of long blonde-wood tables. Some of these tables are clearly set up for lunch, sporting boxes of Girl Scout cookies, jars of Vegemite (someone here is from Australia), and small wire baskets overstuffed with an array of condiments. The others tell a different story; they’re cluttered with monitors, spare robotics parts, tangled black wires, and fully assembled robotic arms that are all trying to master everyday tasks.
During my visit, I see one arm struggling to fold a pair of black pants—not quite a success. Another arm diligently attempts to turn a shirt inside out, showing a kind of determination that indicates it might get there eventually, just not today. Then there’s a third arm, which seems to have found its niche; it’s efficiently peeling a zucchini, successfully depositing the shavings into a separate container.
“Think of it like ChatGPT, but for robots,” Sergey Levine explains, gesturing toward the mechanical activity. Levine, a co-founder and associate professor at UC Berkeley, has the approachable, bespectacled vibe of someone who’s adept at breaking down complex ideas for those who aren’t experts.
Image Credits: Connie Loizos for TechCrunch
What I’m observing is part of a testing loop: data gets gathered at stations here and in various other places—warehouses, homes, anywhere they can set up. This information trains general-purpose robotic foundation models. When a model is trained, it returns to these stations for evaluation. The pants-folder and shirt-turner are experiments in action; the zucchini-peeler is testing the model’s ability to generalize across different vegetables, learning to peel everything from an apple to a potato it has never encountered before.
In addition to this, the company boasts a test kitchen to expose the robots to a variety of environments and challenges. There’s a sophisticated espresso machine that I initially assume is for the staff, but Levine clarifies that it’s actually for the robots to learn from. Those frothed lattes? Just data—not a treat for the engineers, most of whom are absorbed in their computers or their mechanical projects.
The hardware itself is intentionally practical. These arms retail for about $3,500, which Levine describes as “an enormous markup” from the vendor. If they made them in-house, the material costs would drop to under $1,000. A few years back, Levine notes, a roboticist would have been stunned that such arms could perform any tasks at all. But that’s the goal—effective intelligence can make up for subpar hardware.
As Levine steps away, I’m approached by Lachy Groom, who moves through the space with the urgency of someone juggling multiple tasks. At just 31, Groom retains the youthful energy of a Silicon Valley prodigy; he sold his first company after just nine months at 13 years old in Australia (hence the Vegemite).
When I first approached him, he was welcoming a small group of visitors wearing sweatshirts. Initially, his response to my request for a chat was a quick “Absolutely not, I’ve got meetings.” Now he has about ten minutes.
Groom found his calling by following the research coming from Levine’s and Chelsea Finn’s labs. Finn, a former PhD student of Levine’s, now heads her own lab at Stanford focusing on robotic learning. Their names kept popping up in discussions about exciting advancements in robotics. When Groom heard rumors of a potential new startup, he sought out Karol Hausman, a Google DeepMind researcher also involved. “It was one of those meetings where you leave thinking, This is it.”
Groom never set out to be a full-time investor, despite his impressive track record. After working at Stripe as an early employee, he spent around five years as an angel investor, backing companies like Figma, Notion, Ramp, and Lattice while searching for the right opportunity to start or join his own venture. His first investment in robotics, Standard Bots, came in 2021 and reconnected him with a field he loved as a kid constructing Lego Mindstorms. He humorously mentions that he was “on vacation much more as an investor.” But investing was merely a way to stay engaged and connect with people, not the final goal. “I spent five years looking for the right company post-Stripe,” he shares. “Good ideas at the right time, with a good team—that’s rare. It’s all about execution, but you can execute on a bad idea, and it remains a bad idea.”
Image Credits: Connie Loizos for TechCrunch
The two-year-old company has already raised over $1 billion, and when I inquire about its financial runway, Groom quickly clarifies that they don’t actually burn through that much money. Most of their expenditures are on computing power. Moments later, he admits that under the right conditions, he would be open to raising more. “There’s really no limit to how much we can invest,” he notes. “There’s always more compute we can throw at the problem.”
What sets this arrangement apart is Groom’s openness about timelines for making Physical Intelligence profitable. “I don’t provide investors with timelines for commercialization,” he explains, noting that backers like Khosla Ventures, Sequoia Capital, and Thrive Capital have valued the company at $5.6 billion. “It’s odd that they accept that, but they do.” Yet this acceptance may not last, making it prudent for the company to secure ample funding now.
So, what’s the game plan if not immediate commercialization? Quan Vuong, another co-founder from Google DeepMind, outlines their focus on cross-embodiment learning and diverse data sources. If a new hardware platform is developed tomorrow, they won’t need to begin data collection from scratch—they can apply all the knowledge the model already possesses. “The marginal cost of onboarding autonomy to any new robot platform is significantly lower,” he adds.
Currently, the company is collaborating with a select few businesses in various sectors—logistics, grocery, and even a nearby chocolate maker—to assess whether their systems can handle real-world automation. Vuong claims that in some instances, they already are capable. With their “any platform, any task” methodology, the chances for success are broad enough that they can start ticking off tasks ready for automation today.
Physical Intelligence is not the only company striving for this goal. The race to develop general-purpose robotic intelligence—a foundation for more specialized applications akin to the LLM models that took the world by storm three years ago—is intensifying. Pittsburgh-based Skild AI, founded in 2023, recently raised $1.4 billion at a $14 billion valuation and has taken a notably different approach. While Physical Intelligence remains rooted in pure research, Skild AI has already commercially deployed its “omni-bodied” Skild Brain, boasting $30 million in revenue within just a few months last year across sectors like security, warehousing, and manufacturing.
Image Credits: Connie Loizos for TechCrunch
They’ve even publicly challenged competitors, asserting on their blog that many “robotics foundation models” are merely vision-language models “in disguise,” lacking “true physical common sense” due to an over-reliance on pretraining from internet-scale data rather than physics-based simulation and real robotics data.
This indicates a significant philosophical divide. Skild AI believes that commercial deployment will generate a data flywheel that enhances the model with each real-world application, while Physical Intelligence maintains that holding off on immediate commercialization will enable it to create superior general intelligence. Determining who is “more right” will take years.
In the meantime, Physical Intelligence operates with what Groom describes as remarkable clarity. “It’s a pure mission,” he states. “When a researcher has a need, we collect data or find new hardware to support that need. It’s not driven by external forces.” The company originally laid out a 5- to 10-year roadmap for what they thought was feasible, but by month 18, they had already surpassed that timeline.
With about 80 employees, the company has plans to expand, though Groom hopes for slow growth. He admits, “The hardest part is hardware. It’s just really challenging. Everything we’re doing is much tougher than running a software company.” Hardware fails, delivery is often slow, and safety concerns complicate matters.
As Groom hurries off to his next meeting, I find myself watching the robots continue their practice. The pants remain uncooperative, the shirt stays right-side-out, and the zucchini shavings are accumulating nicely.
There are lingering questions—like whether people really want robots peeling vegetables in their kitchens, issues around safety, or concerns about pets reacting to mechanical interlopers at home. Others wonder if the time and resources being invested here are solving significant problems or creating new ones. Critics also question the company’s progress and whether its vision is attainable, contemplating if focusing on general intelligence over specific applications is wise.
If Groom harbors any doubts, he doesn’t let them show. He’s working alongside experts who have tackled these challenges for decades, all believing the timing is finally right—just enough assurance for him.
After all, Silicon Valley has a long history of supporting innovators like Groom, allowing them the freedom to explore and pivot. Even without a clear route to profitability or certainty about how the future market will unfold, the potential for success often outweighs the risks. While not every venture pays off, when they do, the rewards can be significant.
