Physical Intelligence is a San Francisco lab building one robot foundation model, the π series, rather than per-task controllers. Its newest step is the Pi API beta of 13 October 2026, which lets developers fine-tune and host π models for their own robots. Pricing and compliance papers are not published yet.
Physical Intelligence is a robotics lab founded by former Google DeepMind researchers Karol Hausman, Sergey Levine and Chelsea Finn, among others. Its π series of robot foundation models controls several different robot bodies, and the new Pi API beta lets outside developers fine-tune them. Backers include CapitalG, Lux Capital and Thrive Capital.
Founded: 2024 · HQ: San Francisco, CA, USA · Team: 100-250 · CEO: Karol Hausman · Funding: About $1.07B across 3 rounds per press reports (lead investors: CapitalG, Lux Capital, Thrive Capital) · Valuation: $5.6B (Series B, Nov 2025); an $11B round was reported in talks in Mar 2026, close unconfirmed
About Physical Intelligence
Physical Intelligence Inc. was founded in November 2024 in San Francisco by Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail and Quan Vuong. Hausman, Levine, Finn and Ichter came from the robotics group at Google DeepMind, where they worked on robot learning systems such as RT-1 and RT-2, and left to build a single general-purpose robot foundation model instead of narrow, per-task controllers. The company is a Delaware C-Corp with a single San Francisco base.
The product line is the pi (π) series of vision-language-action models, which turn camera images and a language instruction into robot motor commands. π0 arrived on 31 October 2024, and on 4 February 2025 the company released the π0 and π0-FAST weights and code through its openpi repository. π0.5 followed in April 2025 with open-world generalization to unseen kitchens and bedrooms, and π*0.6 in November 2025 added reinforcement learning from real-world experience. π0.7, published on 16 April 2026, is described by the company as a steerable generalist that matches fine-tuned specialists on dexterous tasks and shows first signs of recombining skills for tasks absent from its training data. Research posts in March 2026 added a long and short-term memory method and a fast online reinforcement learning method for precise manipulation.
The biggest recent change is commercial. On 13 October 2026 the company opened the Pi API beta: developers upload data, launch fine-tuning on compute that Physical Intelligence manages, and run the resulting models through a hosted inference gateway on their own robots. The announcement names early partners including Actor, which deploys autonomous heavy machinery, and Almond, whose dual-arm Axol robot ran a model obtained after four days of iteration. It says most showcased deployments used cloud-hosted inference and that network connectivity caused problems for some. Access is by application ("tell us what you want to build"), and the post publishes no pricing.
Funding is reported rather than disclosed in detail. Press coverage puts the total at about $1.07 billion across three rounds, including a $600 million Series B in November 2025 at a $5.6 billion valuation led by Alphabet's CapitalG, with Lux Capital, Thrive Capital, Index Ventures, T. Rowe Price, NVentures, Amazon and Jeff Bezos named as participants. Bloomberg reported in March 2026 that the company was discussing about $1 billion at a valuation above $11 billion. We found no company announcement confirming that round closed, and aggregator listings disagree, so treat the $5.6 billion figure as the last confirmed valuation. Headcount estimates range widely by data provider, so the 100-250 team size here is a band, not a count.
Where the alternatives win is worth stating plainly. Skild AI is also building a robot-agnostic foundation model. Gemini Robotics from Google DeepMind brings Gemini-scale multimodal reasoning into robot control. Isaac GR00T from NVIDIA offers open, customizable humanoid models backed by its simulation and compute stack. Full-stack makers such as Figure AI tune models to their own hardware, which can mean tighter integration than a hardware-agnostic model offers. Physical Intelligence owns no robot hardware and depends on partners to put its models in the field.
Teams that need published pricing, a service-level agreement or audit reports should wait. We found no SOC 2, ISO 27001 or HIPAA attestation and no trust center on the sites read, and no EU AI Act classification or data-processing terms for the new API. Treat that as an absence of published evidence, not proof that none exists, and ask the company directly before sending production data. For the general-purpose labs that sit beside robotics specialists, see OpenAI.
Mission
Develop a foundation model that can control any robot to do any task.
Products
- pi-zero (π0) and π0-FAST (Open weights and code (openpi, Feb 2025)): https://github.com/Physical-Intelligence/openpi
- pi-0.5 (π0.5) (Generalist VLA policy with open-world generalization): https://www.pi.website/blog/pi05
- pi*-0.6 (π*0.6) (Generalist VLA policy trained with RL from experience): https://www.pi.website/blog/pistar06
- pi-0.7 (π0.7) (Steerable generalist VLA policy, cross-embodiment): https://www.pi.website/blog/pi07
- Pi API (beta) (Fine-tuning and hosted inference for π models, by application): https://www.pi.website/blog/api-beta
Links
Frequently Asked Questions
What does Physical Intelligence sell, and is there an API?
Physical Intelligence builds the π family of robot foundation models, and since October it offers them through the Pi API beta. Developers upload demonstration data, fine-tune a π model on infrastructure the company manages, and run it on their own robots through a hosted gateway. Access is by application and the announcement lists no prices. The company does not sell robots, a consumer app or a general chat product.
Who founded Physical Intelligence and who leads it?
Seven co-founders started the company in late 2024, among them Lachy Groom, Adnan Esmail and Quan Vuong beside the four former Google DeepMind researchers. Karol Hausman is chief executive. Sergey Levine and Chelsea Finn, professors at UC Berkeley and Stanford, act as chief scientists. We found no announced leadership change since founding.
How much money has Physical Intelligence raised?
Press reports put total funding near $1.07 billion across three rounds. The latest confirmed round is a Series B of $600 million that closed in late 2025 with CapitalG as lead and a $5.6 billion price tag on the company. Bloomberg reported in March 2026 on talks for a further round of roughly $1 billion that would value the company above $11 billion, but no company announcement we could read confirms a close. Revenue is undisclosed and the company is privately held.
Are the π models open source?
Only some of them. The openpi repository on GitHub holds the earliest π0 release and its faster FAST variant, published in February 2025, so researchers can run and adapt them locally. Newer models, including the April 2026 release, are reached through the Pi API beta rather than as downloads. Licence terms sit in the repository and papers, so read them before commercial use.
How does Physical Intelligence compare with Skild AI, Google DeepMind and NVIDIA?
Skild AI is the closest like-for-like rival, also pursuing a robot-agnostic foundation model. Google DeepMind draws on large multimodal models and far greater resources, and NVIDIA pairs open humanoid models with its simulation and compute tools. Figure AI trains models for its own humanoid, which can integrate more tightly than a hardware-agnostic approach. Physical Intelligence offers one base model fine-tuned per robot through an API, but owns no hardware and publishes no pricing or compliance papers yet.
Top Alternatives
- Figure AI: Pick π for hardware-agnostic foundation model; pick Figure for vertically integrated humanoid.