PHYSICAL INTELLIGENCE

The founding team reads like an all-star roster from the robotics research world. Sergey Levine runs a famous lab at UC Berkeley.

Chelsea Finn is a Stanford professor who created a landmark method for teaching machines to learn quickly. Karol Hausman worked on robot foundation models at Google DeepMind.

With Brian Ichter, investor Lachy Groom, Adnan Esmail and Quan Vuong, they teamed up in 2024 with one shared frustration. Robots were still being programmed task by task, factory by factory.

Every new job meant starting from scratch. They wanted one general model that could drive any robot, the way one language model can write about anything.

So they raised a giant seed and got to work.

There is no revenue yet. The plan is to build a single AI foundation model for robots and license it out.

Think of it like an operating system for machines with arms. A warehouse company, a home-robot maker, or a factory would plug into the model instead of writing their own robot software from zero.

Physical Intelligence sells the intelligence. Other companies build the bodies.

If it works, they become the layer every robot runs on. If it does not, they are an expensive research lab.

Pi-zero is the flagship. It is a general AI model that turns a camera view and an instruction into robot movements.

The point is that it is not tied to one machine. The team has demoed it folding laundry, bussing tables, and packing boxes across several robot arms.

It is a research model for now, not a shipping product. Its 2025 follow-up, pi-0.5, focused on working in homes it had never seen before.

The next models are meant to be smarter and more reliable versions of the same idea.

The move here is speed and proof. Seven months after the seed, the team published demos of its model, called pi-zero.

It folded laundry, assembled cardboard boxes, and did warehouse tasks. The clever part is that the same model ran across different robot bodies.

That is the whole pitch. One brain, many machines.

Instead of quietly researching for years, they showed working video early. The videos did the fundraising.

The $400 million round followed fast.

Laundry-folding demos are not a business. The hard part is going from a cool video to something reliable enough to sell.

Robots that work 95 percent of the time are useless in a warehouse that needs 99.9 percent. The company is also burning cash with no product and no revenue.

It has raised at $2.4 billion and then $5.6 billion before proving anyone will pay. Its robots are only in early trials, such as folding laundry in short-term rentals and boxes at a San Francisco chocolate maker.

Rivals like Figure and Skild AI are chasing the same prize. The clock is the real enemy here.

The backer list is the story. The $70 million seed in March 2024 was led by Thrive Capital, with Khosla Ventures, Lux Capital, OpenAI and Sequoia Capital piling in.

The November 2024 round of $400 million at a $2.4 billion valuation brought in Jeff Bezos personally, plus Thrive and Lux again, with OpenAI, Redpoint Ventures and Bond joining. In November 2025 Alphabet's growth fund CapitalG led a $600 million round at $5.6 billion, with Index Ventures and T.

Rowe Price among the new names. By 2026 Bloomberg reported it was raising about $1 billion more at a valuation around $11 billion.

When Amazon's founder, the maker of ChatGPT and Google's own fund all back your robot brain, people notice.