Upon arriving at the headquarters of Physical Intelligence in San Francisco, the only noticeable sign is a uniquely colored pi symbol on the door. Inside, the atmosphere is bustling, devoid of a formal reception area or flashy logos.
The interior resembles a spacious concrete box, softened by a mix of long blonde-wood tables. Some tables are set up for meals, featuring Girl Scout cookie boxes, jars of Vegemite, and assorted condiments, while others are cluttered with monitors, spare robotic parts, and robotic arms engaged in various tasks.
During my visit, I observed one robotic arm struggling to fold a pair of black pants, while another was determinedly trying to turn a shirt inside out. A third arm, however, was efficiently peeling a zucchini, successfully depositing the shavings into a designated container.
"Imagine it as ChatGPT for robots," explains Sergey Levine, a co-founder and associate professor at UC Berkeley, as he gestures towards the mechanical choreography happening in the room.
This testing phase involves gathering data from robot stations in various settings--warehouses, homes, and more--to train general-purpose robotic foundation models. Each new model undergoes evaluation at these stations, with the pants-folder and shirt-turner serving as experimental subjects. The zucchini-peeler is assessing whether the model can adapt its peeling technique to different vegetables.
The company operates test kitchens in this facility and other locations, including private homes, utilizing readily available hardware to expose robots to diverse environments and challenges. A sophisticated espresso machine sits nearby, not for the staff, but for the robots to learn from, as any lattes made are merely data for the engineers present.
The robotic arms, priced around $3,500, are intentionally unassuming, as Levine notes that manufacturing them in-house could reduce costs significantly. He emphasizes that the intelligence of the robots can compensate for the limitations of the hardware.
As I continue my conversation with Lachy Groom, the co-founder who exudes the energy of someone managing multiple tasks, he shares insights into his journey. Groom, who gained recognition in Silicon Valley at a young age, initially sold his first company at just 13 years old in Australia.
While Groom originally did not plan to become a full-time investor, he spent years investing in startups like Figma and Notion after leaving Stripe. His first foray into robotics investment came in 2021, reigniting a passion he had since childhood. He humorously reflects on how investing allowed him more leisure time, but it was merely a stepping stone until he found the right company to join.
The startup has successfully raised over $1 billion, and Groom clarifies that their spending is primarily on computational resources. He expresses a willingness to raise more funds under favorable conditions, highlighting the limitless potential for computational resources to tackle challenges.
What sets Physical Intelligence apart is Groom's unconventional approach to investor relations, as he does not provide a timeline for commercialization. This transparency seems to resonate with backers, allowing the team to focus on long-term goals without external pressures.
Co-founder Quan Vuong elaborates on their strategy, emphasizing cross-embodiment learning and diverse data sources, enabling the company to adapt quickly to new hardware platforms without starting data collection from scratch.
Physical Intelligence is collaborating with various companies across different sectors to evaluate their systems for real-world automation. Vuong asserts that in certain cases, their technology is already viable for practical applications.
The pursuit of general-purpose robotic intelligence is intensifying, with other startups like Skild AI also making strides in the field. While Physical Intelligence focuses on research, Skild AI is already deploying its technology commercially, generating significant revenue within a short timeframe.
Despite differing philosophies regarding commercialization, both companies are contributing to the advancement of robotic intelligence. Groom remains confident in the team's capabilities and the timing of their innovations, as they navigate the complexities of hardware development.
As I observe the robots continue their tasks, I am left pondering the future of domestic robotics and the broader implications of their integration into daily life. With a clear vision and a dedicated team, Physical Intelligence is poised to redefine the landscape of robotics.