Physical Intelligence raises $600 million at $5.6 billion as CapitalG leads bet on a shared robot brain

Physical Intelligence, the San Francisco start-up building general-purpose AI models for robots, has raised $600 million at a $5.6 billion valuation, according to its investor Outset Capital. Alphabet's growth fund CapitalG led the round and confirmed its lead role in a note published on November 20, 2025.

Outset Capital, which describes Physical Intelligence as a portfolio company, said Lux Capital, Thrive Capital, Jeff Bezos, Index Ventures and T. Rowe Price took part alongside CapitalG. CapitalG called the financing a Series B. Neither investor disclosed how the $600 million splits between new and existing backers, whether the $5.6 billion figure is measured before or after the new money, or what the company plans to spend it on line by line.

The raise lands three days after Physical Intelligence published π*0.6 on November 17, 2025, a version of its vision-language-action model trained further with reinforcement learning on real robots. In the accompanying paper, posted to arXiv on November 18, the company's researchers describe a method called RECAP that feeds demonstrations, data the robot gathers while working and corrections from human teleoperators into one self-improvement loop. On the hardest tasks they tested, the authors say RECAP more than doubles task throughput and roughly halves the failure rate.

CapitalG's partners Jill Chase and Manmeet Gujral wrote that π*0.6 runs for hours making espresso drinks, folds unfamiliar laundry in homes it has never seen and assembles packaging boxes in working factories, with success rates above 90%. Outset Capital added that the model is already deployed commercially at a laundromat in San Francisco. Neither investor named the operator or gave fleet sizes.

The company sells no robot of its own. Its bet is on the software layer, a single policy that can be trained once on data from many machines and then run on different arms and grippers. It published its first generalist policy, π0, on October 31, 2024, and released π0 as open source on February 4, 2025. CapitalG's note names Karol Hausman as co-founder and chief executive and Lachy Groom as co-founder, and singles out the researchers Sergey Levine, Chelsea Finn, Brian Ichter and Quan Vuong and engineering leader Adnan Esmail.

The round puts Physical Intelligence among the most richly funded robot software companies in the United States, though still well behind the humanoid makers that build their own hardware. Figure AI closed a Series C of more than $1 billion at a $39 billion post-money valuation on September 16, 2025, about seven times the value investors put on Physical Intelligence. Skild AI, its closest rival in robot foundation models, last disclosed a $300 million Series A at $1.5 billion in July 2024. FieldAI announced $405 million across two rounds in August 2025, and Dyna Robotics raised a $120 million Series A in September.

The technology at stake is easy to describe and hard to make work. A vision-language-action model takes camera images and a written instruction and outputs motor commands. Training on demonstrations alone gives a robot that copies people but stalls when something unexpected happens. Reinforcement learning on deployed robots lets the model improve from its own successes and mistakes, which is what π*0.6 claims to do at a scale useful to a business. Throughput matters because a robot that is reliable but slow does not pay back its cost on a factory floor or in a laundromat.

For industrial buyers in the US and Europe, the signal is that capital is flowing to whoever can make generic robot arms productive without months of custom programming. That threatens the integrator model that dominates automation sales today and favours hardware makers willing to open their machines to an outside brain. It also raises the price of talent, since a handful of research groups now compete for the same few hundred people who have trained these models on real robots.

The risks are plain. Physical Intelligence has disclosed no revenue, no customer count and no robot fleet numbers. Its strongest performance claims come from its own paper and from investors with a stake in the outcome. Model companies also depend on hardware partners for data and deployment, and a humanoid maker with its own model, such as Figure, can keep that data in house.