Physical Intelligence: $5.6B Robot AI Startup 2026
Physical Intelligence is a 2024-founded San Francisco robotics startup building pi-zero VLA robot foundation models, valued at $5.6 billion in late 2025.
Physical Intelligence, founded in 2024 in San Francisco by ex-Google DeepMind researchers including Karol Hausman, Sergey Levine, and Chelsea Finn, builds the pi-zero (π0) line of vision-language-action foundation models for robots. Backed by Jeff Bezos, Amazon, CapitalG, Thrive Capital, and NVIDIA, it has raised over $1 billion and was valued at $5.6 billion in late 2025, with talks of an $11 billion round in 2026. Its models already power real pilot deployments such as folding laundry and packing chocolate boxes.
Physical Intelligence is a robotics AI startup founded in November 2024 in San Francisco by Karol Hausman, Sergey Levine, and Chelsea Finn, all formerly of Google DeepMind robotics research. It has raised over $1 billion across three rounds and was valued at $5.6 billion in November 2025. Its flagship product is the pi-zero (π0) family of vision-language-action foundation models, which let a single AI brain control many different robot bodies on real tasks like folding laundry and packing boxes.
Founded: 2024 · HQ: San Francisco, CA, USA · Team: 100-250 · CEO: Karol Hausman · Funding: $1.07B total across 3 rounds (lead investors: CapitalG, Lux Capital, Thrive Capital) · Valuation: $5.6B (Series B, Nov 2025); in talks for $11B round as of mid-2026
About Physical Intelligence
Physical Intelligence Inc. was founded in November 2024 in San Francisco, California by Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail and Quan Vuong. Hausman, Levine, Finn and Ichter came from Google DeepMind's robotics research group, where they worked on large-scale robot learning systems such as RT-1 and RT-2 before concluding the field needed a dedicated company focused on a single general-purpose robot foundation model rather than narrow, per-task systems. Incorporated as a Delaware C-Corp and based in San Francisco, the company assembled a small but dense team of roboticists and machine learning researchers drawn from Stanford, UC Berkeley and Google. The company's flagship product line is the pi (π) series of vision-language-action (VLA) models. π0, unveiled in late 2024 and partly open-sourced through the openpi GitHub repository in early 2025, was the first generalist policy able to control multiple distinct robot embodiments, including arms from Trossen and ARX and mobile manipulators, on tasks such as folding laundry, bussing tables, assembling cardboard boxes and making coffee, all from a single set of model weights. The π0-FAST variant added an autoregressive tokenization scheme for faster inference, while π0.5 (April 2025) extended the approach with open-world generalization to new homes and objects never seen in training. Recent releases have come quickly. π0.6, announced November 17, 2025 with a published model card, introduced reinforcement learning from real-world experience, improving both task success rate and throughput over π0.5 on the same hardware. In April 2026 the company unveiled π0.7, which it says reaches specialist-level dexterous performance without task-specific fine-tuning, generalizes across robot embodiments and environments, and can recombine learned skills to handle tasks it was never explicitly trained on, such as cleaning an unfamiliar kitchen from a single spoken instruction. Physical Intelligence has raised a total of approximately $1.07 billion across three rounds from 14 investors. An initial round of roughly $70 million accompanied the π0 launch in late 2024. A Series A of $400 million followed in 2025 at a $2.4 billion valuation. A Series B of $600 million closed around November 2025 at a $5.6 billion valuation, led by Alphabet's growth fund CapitalG with participation from Lux Capital, Thrive Capital, Index Ventures, T. Rowe Price, NVIDIA's NVentures, Amazon, Jeff Bezos and Klarna co-founder Sebastian Siemiatkowski. As of mid-2026 the company is reportedly in talks to raise roughly $1 billion more at an $11 billion valuation, with Founders Fund and Lightspeed Venture Partners named as participants. The company is pre-revenue in the traditional sense and has not disclosed pricing or an API product. Its near-term business model centers on real-world pilot deployments in partnership with robot hardware makers and operators: its models have been used to fold laundry in short-term rental units in San Francisco and to fold cardboard packaging boxes in the back rooms of Dandelion Chocolate. The openpi GitHub repository makes base π0 and π0.5 weights available for research use, which the company uses to build developer mindshare while reserving its newest models, π0.6 and π0.7, for commercial partners. CEO and co-founder Karol Hausman leads the company alongside co-founders and chief scientists Sergey Levine and Chelsea Finn, both robotics and machine learning professors at UC Berkeley and Stanford respectively, and Brian Ichter, formerly of Google Brain's robotics team. Headcount estimates vary by source, with Tracxn citing roughly 209 employees as of February 2026 versus an earlier PitchBook figure near 80, reflecting the company's rapid hiring through 2025 and 2026. The team is concentrated in a single San Francisco office, with engineering, robotics hardware integration and research functions co-located. The company's stated mission is to develop a foundation model that can control any robot to do any task, treating robotics the way large language model labs treat text: one base model, fine-tuned or prompted for many downstream uses, rather than bespoke controllers per robot and task. Its research agenda centers on vision-language-action architectures, flow-matching action heads for continuous control, cross-embodiment transfer between robot arms and mobile platforms, and reinforcement learning from real-world deployment data, with findings published on arXiv and in model cards such as the November 2025 π0.6 release notes. Among pure foundation-model labs, Physical Intelligence's closest rivals are Skild AI, which is building a similarly robot-agnostic 'Skild Brain', and Google DeepMind's Gemini Robotics line, which brings web-scale multimodal reasoning into robot control. NVIDIA's Isaac GR00T competes directly on open, customizable humanoid foundation models. Physical Intelligence's advantage is being hardware-agnostic with models already running in paid real-world pilots rather than lab demos; its main weakness is that it owns no robot hardware or distribution channel of its own and depends on hardware partners such as Trossen, ARX and pilot-site operators to get its models into the field. No EU AI Act classification, GDPR commitments or formal compliance certifications have been publicly disclosed as of mid-2026; as a US-based robotics research company without a public API product, it currently falls outside the consumer-facing data protection disclosures typical of SaaS AI vendors. Looking ahead, the reported $1 billion round at an $11 billion valuation, if it closes, would roughly double the company's valuation within seven months of its Series B, underscoring investor expectations that pi-zero-class models will be the operating system for the next wave of humanoid and mobile-manipulator robots entering factories and homes through 2026 and 2027.
Mission
Develop a foundation model that can control any robot to do any task.
Products
- pi-zero (π0) (Hosted research model / open weights): https://github.com/Physical-Intelligence/openpi
- pi-0.5 (π0.5) (Hosted research model / open weights): https://www.physicalintelligence.company/blog/pi05
- pi-0.6 (π0.6) (Generalist VLA policy with RL from experience): https://www.physicalintelligence.company/blog
- pi-0.7 (π0.7) (Generalist VLA policy, cross-embodiment): https://www.physicalintelligence.company/blog
Links
Website · GitHub · Twitter · LinkedIn · Blog · Docs
Frequently Asked Questions
What is Physical Intelligence and what do they build?
Physical Intelligence is a robotics AI company founded in November 2024 in San Francisco by a team of former Google DeepMind robotics researchers and Stanford and UC Berkeley professors. The company builds the pi (π) series of vision-language-action models, which are foundation models that take in camera images and natural language instructions and output low-level robot motor commands. Its flagship release, pi-zero (π0), was the first generalist policy able to control several different robot arms and mobile manipulators from a single set of weights, performing tasks such as folding laundry, bussing tables, assembling boxes, and making coffee. Later versions, π0.5, π0.6, and π0.7, added open-world generalization, reinforcement learning from real-world experience, and the ability to recombine learned skills for tasks the model was never explicitly trained on. Base versions of these models are open-sourced through the openpi GitHub repository for researchers, while the newest versions power commercial pilot deployments. Physical Intelligence does not sell a public consumer app or API; instead it partners with robot hardware makers and operators to deploy its models in real settings. Within robotics AI, it is considered one of the leading 'foundation model for any robot' labs alongside Google DeepMind and Skild AI.
Who founded Physical Intelligence and who is the CEO?
Physical Intelligence was founded in November 2024 by seven people: Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail, and Quan Vuong. Karol Hausman serves as Chief Executive Officer and co-founder. Sergey Levine, a professor at UC Berkeley, and Chelsea Finn, a professor at Stanford, serve as chief scientists, bringing decades of combined academic research in robot learning and reinforcement learning. Hausman, Levine, Finn, and Ichter all previously worked on large-scale robot learning at Google DeepMind, including projects like RT-1 and RT-2, and left after concluding that a focused company building one general-purpose robot foundation model could move faster than a large research organization with many competing priorities. The company was incorporated as a Delaware C-Corp with its headquarters in San Francisco. There have been no publicly reported leadership changes since founding, and the founding team remains in place as of mid-2026.
How much funding has Physical Intelligence raised?
Physical Intelligence has raised approximately $1.07 billion across three disclosed funding rounds from 14 investors. The company's first round, around $70 million, closed alongside the launch of its π0 model in late 2024. A Series A round of $400 million followed in 2025 at a $2.4 billion valuation. A Series B round of $600 million closed around November 2025 at a $5.6 billion valuation, led by Alphabet's growth-stage fund CapitalG, with participation from Lux Capital, Thrive Capital, Index Ventures, T. Rowe Price, NVIDIA's venture arm NVentures, Amazon, Jeff Bezos, and Klarna co-founder Sebastian Siemiatkowski. As of mid-2026, the company is reportedly in talks to raise roughly $1 billion more at an $11 billion valuation, with Founders Fund and Lightspeed Venture Partners named as prospective participants. The company has not disclosed revenue or an IPO timeline; it remains privately held.
What products does Physical Intelligence make?
Physical Intelligence's products are all part of the pi (π) family of vision-language-action foundation models for robots. Pi-zero (π0), released in late 2024, was the first model in the line, capable of controlling multiple robot embodiments including arm-based manipulators on household and light industrial tasks. Pi-0.5, released April 2025, added open-world generalization so robots could handle unfamiliar rooms and objects. Pi-0.6, announced November 2025 with a published model card, introduced reinforcement learning from real-world deployment experience to improve task success rates and speed. Pi-0.7, unveiled in April 2026, reaches specialist-level dexterity without task-specific fine-tuning and can recombine learned skills for novel tasks. Base versions of π0 and π0.5 are open-sourced through the openpi GitHub repository for researchers and developers at no cost, while π0.6 and π0.7 are reserved for commercial pilot partners. The company does not currently sell a standalone consumer app, developer API, or enterprise SaaS product; its models are deployed through hardware and operator partnerships.
Where is Physical Intelligence headquartered and how big is the team?
Physical Intelligence is headquartered in San Francisco, California, where its research, engineering, and robotics hardware integration teams are co-located in a single office. The company does not maintain publicly disclosed satellite offices in other cities or countries as of mid-2026. Headcount estimates vary by data provider: Tracxn reported approximately 209 employees as of February 2026, while an earlier PitchBook estimate put the figure closer to 80, reflecting rapid hiring through 2025 and into 2026 as the company scaled from its Series A to Series B and prepared for a possible third major round. The team is described as research-dense, drawing heavily from Stanford, UC Berkeley, and Google DeepMind robotics alumni. No layoffs or restructuring have been publicly reported.
What is Physical Intelligence's mission or research focus?
Physical Intelligence's stated mission is to develop a foundation model that can control any robot to do any task, applying the same 'one base model, many downstream uses' approach that large language models brought to text to the physical world of robotics. In concrete terms, this means training a single vision-language-action model that can be deployed across different robot arms, mobile manipulators, and environments without building a separate controller for each robot and task. The company's active research areas include vision-language-action model architectures, flow-matching action heads for smooth continuous control, cross-embodiment transfer so skills learned on one robot generalize to another, and reinforcement learning from real-world experience collected during pilot deployments. Findings are published as arXiv papers, including the π0 and π0.5 papers, and as detailed model cards such as the November 2025 π0.6 release. Unlike some AI labs, Physical Intelligence has not published a formal responsible scaling policy, reflecting its earlier stage and narrower focus on robotics rather than general-purpose AI safety.
Is Physical Intelligence compliant with SOC 2, GDPR, HIPAA?
As of mid-2026, Physical Intelligence has not publicly disclosed any formal compliance certifications such as SOC 2, ISO 27001, or HIPAA, and does not maintain a public trust center page. This is consistent with its current stage and business model: the company does not sell a hosted consumer API or SaaS product that would typically require these certifications for enterprise procurement. Its commercial activity to date consists of research pilots and partnerships with robot hardware makers and site operators, where data handling terms are negotiated directly with each partner rather than governed by a standardized public policy. No GDPR-specific data processing addendum, EU data residency commitment, or EU AI Act risk classification has been published. As the company moves toward broader commercial deployment following its π0.7 release and a possible new funding round, formal compliance documentation would be expected to follow, but as of mid-2026 none has been confirmed.
Who are Physical Intelligence's main competitors?
Physical Intelligence's closest competitor among pure foundation-model labs is Skild AI, which is also building a robot-agnostic foundation model, branded the 'Skild Brain', intended to run across robot hardware it has never seen. Google DeepMind competes through its Gemini Robotics line, which brings the web-scale multimodal reasoning of the Gemini models into robot control, an advantage Physical Intelligence cannot match given its smaller scale, though Physical Intelligence counters with models already running in real paid pilots rather than primarily lab demonstrations. NVIDIA's Isaac GR00T competes on open, customizable humanoid foundation models backed by NVIDIA's simulation and compute ecosystem, an infrastructure advantage Physical Intelligence lacks. Full-stack humanoid makers such as Figure AI train competitive in-house models like Helix tuned specifically to their own hardware, giving them tighter hardware-software integration than Physical Intelligence's hardware-agnostic approach can offer. An emerging risk not on the radar 12 months ago is well-funded Chinese humanoid and robotics AI labs scaling VLA research at speed.