Physical Intelligence: Who They Are, Pi0, and Funding

What Is Physical Intelligence?

Physical Intelligence is a robotics AI startup founded in 2024 and based in San Francisco. Unlike Tesla, Unitree, or Xiaomi, it does not primarily build its own robot hardware — its product is a general-purpose "foundation model" for robot control, the same idea as a large language model but trained to output physical actions instead of text.

The company's stated goal is a single model that can be pointed at many different robot bodies and tasks — folding laundry, bussing a table, packing a box — using natural-language instructions, the way a chatbot follows a text prompt. That is the core pitch behind embodied AI as a category; see our embodied AI explainer for the underlying concept.

This page covers who founded Physical Intelligence, what its Pi0 and Pi0.5 models actually do, its funding and valuation history, and where it fits next to open robotics projects like LeRobot.

Founders And Team

Physical Intelligence was founded by a group with deep roots in robotics and AI research: Karol Hausman (CEO, previously at Google DeepMind), Sergey Levine (Chief Scientist, a UC Berkeley professor known for robot learning research), Chelsea Finn (a Stanford professor working on robot learning), Brian Ichter (previously Google Research), and Adnan Esmail.

That research pedigree is part of why the company attracted investment so quickly after founding — the team combined academic robot-learning credibility with hands-on experience at Google's robotics and AI groups.

Reporting on the founders and team is still developing as the company scales past its early research-lab footprint; treat headcount and org details as changing quickly, and verify current leadership on the company's own site before citing it elsewhere.

Pi0 And Pi0.5: What The Models Do

Pi0 ("pi-zero") is Physical Intelligence's flagship robot foundation model, first shown in October 2024 and open-sourced in February 2025 under the OpenPI project. It is built on a vision-language-action architecture: it takes in camera images and a natural-language instruction and outputs the motor commands to control a robot arm or mobile manipulator.

In demonstrations, Pi0 has been shown folding laundry, bussing tables, assembling boxes, and handling other everyday manipulation tasks across more than one robot platform — the point being generalization across bodies and tasks rather than a single robot trained for one job.

Pi0.5 is the follow-up model, aimed at better open-world generalization — performing tasks in home environments the model was not specifically trained on. It is best understood as an iteration on the same vision-language-action recipe, with more training diversity rather than a fundamentally different approach.

Because Pi0 is open-sourced, it has become a reference point for the broader open robot-learning community, including projects like Hugging Face's LeRobot — see our LeRobot how-to guide for how to actually run open robot-policy code yourself.

• Pi0 — released Oct 2024, open-sourced Feb 2025 (OpenPI).
• Vision-language-action architecture: camera + text instruction → robot motor commands.
• Demonstrated tasks: laundry folding, table bussing, box assembly, general manipulation.
• Pi0.5 — follow-up model focused on open-world / home-environment generalization.

Funding And Valuation

Physical Intelligence has raised more than $1.1 billion across three funding rounds since its 2024 founding, an unusually fast fundraising pace even by AI-startup standards. A $600 million round in November 2025 valued the company at approximately $5.6 billion, led by Alphabet's independent growth fund CapitalG, with backers including Thrive Capital and Lux Capital.

As of mid-2026, reporting indicates Physical Intelligence is in talks for a new round of roughly $1 billion that could push its valuation above $11 billion, with Founders Fund and Lightspeed Venture Partners named as potential participants alongside existing investors.

Figures like these move quickly and are frequently revised as rounds close — treat any specific number as a snapshot at the cited date, and verify against a current source before quoting it elsewhere.

Physical Intelligence Vs. Robot Hardware Makers

The easiest way to place Physical Intelligence is by what it is not: it is not a competitor to Unitree, Tesla Optimus, or Xiaomi CyberOne in the sense of building a robot body to sell or deploy. Those companies build the hardware; Physical Intelligence's bet is that the software "brain" running on many hardware platforms is the more defensible, more scalable layer.

That makes it more comparable to foundation-model labs like OpenAI or Anthropic than to a robotics hardware manufacturer — except trained on physical action data instead of, or in addition to, text. See our robotics news hub for how this software-first approach compares to hardware-first humanoid strategies. Rivals in the same policy-not-hardware category include Skild AI and, on the narrower industrial-cobot side of the field, Generalist AI, whose GEN-1 model targets existing cobot arms rather than a general-purpose humanoid body.

The open question for the category is the same one every foundation-model company faces: whether a general-purpose model actually outperforms task-specific training on the robots it is deployed on, at the reliability level real-world deployment requires.