What is Physical Intelligence?
Foundation models that give any robot a general-purpose brain
- Category
- Robotics AI / Foundation Models
- Headquarters
- San Francisco, CA (396 Treat Ave)
- Founded
- 2024
- Employees
- ~200 (Feb 2026)
- Total Funding
- $1.07B+ confirmed across 3 rounds
- Latest Valuation
- $5.6B (Series B, Nov 2025); ~$11B reported (Series C in talks, Mar 2026)
What is Physical Intelligence?
Physical Intelligence (stylized π) is a San Francisco-based AI company that builds Vision-Language-Action (VLA) foundation models — software brains that let any robot understand natural language commands and execute complex physical tasks, without the manufacturer having to train it from scratch.
Founded in early 2024 by seven researchers and operators from Google DeepMind, UC Berkeley, Stanford, and Anduril Industries, Physical Intelligence is building what it calls the "physical intelligence layer" for the robotics industry: a single generalist policy that can control different robot platforms across different environments and tasks. Its flagship π0 family of models has been demonstrated folding laundry, assembling cardboard boxes, peeling vegetables, sorting warehouse packages, and — most recently — completing novel tasks it was never explicitly trained on (π0.7, released April 2026).
Physical Intelligence does not manufacture hardware. It is a pure-software company pursuing an "Android for robots" model: foundation models licensed as general-purpose intelligence that sit on top of third-party robotic bodies. Early commercial deployments include robots running in a real San Francisco laundromat (via partner Weave Robotics and their Isaac robot) and warehouse order-packaging lines (via partner Ultra). The company open-sourced its π0 base weights and training code on GitHub in February 2025 (Physical-Intelligence/openpi), and added PyTorch support alongside JAX in September 2025, dramatically broadening adoption within the global robotics research community.
Revenue is nascent but growing through enterprise licensing agreements and a private-beta fine-tuning API; the company's primary expenditure remains compute infrastructure for training and iterating its large VLA models. As of June 2026, Physical Intelligence has raised approximately $1.07 billion across three rounds at a $5.6 billion Series B post-money valuation, with reported advanced talks for a ~$1 billion Series C at an $11 billion valuation. With roughly 200 employees — up from approximately 80 in early 2025 — it is among the most well-capitalized and technically prestigious robotics-AI startups in the world.
What does Physical Intelligence offer?
Physical Intelligence's product portfolio centers on generalist VLA foundation models and supporting tooling for the robotics ecosystem, spanning open-source research releases and commercial enterprise access.
- π0 Generalist Policy· Foundation Model
- π0.5 Open-World Generalization (Sep 2025)· Foundation Model
- π*0.6 RL-Trained Model (Nov 2025)· Foundation Model
- π0.7 Steerable Compositional Model (Apr 2026)· Foundation Model
- FAST Action Tokenization· Research / Tooling
- Fine-Tuning API (private beta)· Enterprise Access
- openpi Open-Source Repository· Open Source
- Partner Data Pre-Training Program· Enterprise Access
- Vision-Language-Action (VLA) Architecture· Core Technology
- Cross-Embodiment Learning· Core Technology
- Flow Matching / Diffusion Action Head· Core Technology
How does Physical Intelligence make money?
Physical Intelligence pursues a B2B software licensing model, monetizing its VLA foundation models through a private-beta fine-tuning API and enterprise partnership agreements with robotics companies that want general-purpose AI embedded in their hardware fleets.
The core go-to-market is a foundation-model-as-a-service model for robotics OEMs and automation companies. Partners such as Weave Robotics (home laundry, SF Bay Area) and Ultra (warehouse e-commerce packaging) integrate Physical Intelligence's π models into their robot fleets. The company earns licensing revenue and, in exchange, receives partner data that flows back into pre-training — improving model performance across all customers. Public pricing has not been disclosed; the company has not confirmed third-party estimates for per-robot subscription pricing. Enterprise agreements likely include access to the fine-tuning API, data-sharing arrangements, and on-premises or edge deployment options for latency-sensitive applications.
A parallel channel is the open-source community flywheel. Physical Intelligence released π0 weights and code on GitHub in February 2025, and added PyTorch support in September 2025, driving broad academic adoption. The openpi repository supports fine-tuning on the ALOHA platform, the DROID dataset, and the Libero benchmark, giving researchers easy entry points. The company earns no direct revenue from open-source usage but treats it as top-of-funnel demand generation and recruiting — the same playbook used by frontier LLM providers like Mistral.
Revenue remains early-stage; Physical Intelligence is effectively pre-commercial at scale, prioritizing model capability and hardware-agnostic generalization over near-term monetization. Growth is expected to follow deployment density: as robot fleets running π models scale from dozens to hundreds of units per partner, subscription and licensing revenue scales with them. The long-term monetization thesis is a percentage take-rate on robot intelligence across a global robotics market projected to reach $280 billion by 2034.
Who leads Physical Intelligence?
Physical Intelligence was co-founded in 2024 by seven researchers and operators from Google DeepMind, UC Berkeley, Stanford, Anduril Industries, and Stripe. CEO Karol Hausman leads the company day-to-day alongside a founding team that spans elite academic research and commercial operations experience.
- Karol HausmanCo-Founder & CEO2024–presentFormer Staff Research Scientist at Google DeepMind and adjunct professor at Stanford; PhD from USC; specializes in robot manipulation learning and cross-embodiment policies. Also serves as Secretary and CFO per California corporate filings.
- Sergey LevineCo-Founder & Chief Scientist2024–presentUC Berkeley professor and one of the most-cited researchers in robot learning globally; his lab's work on deep reinforcement learning and imitation learning forms the academic backbone of Physical Intelligence's models.
- Chelsea FinnCo-Founder & Research Lead2024–presentStanford associate professor known for meta-learning and sim-to-real transfer; on leave while co-shaping PI's research roadmap; her MAML algorithm underpins cross-task generalization approaches used in π models.
- Brian IchterCo-Founder & VP Engineering2024–presentFormer Google Brain and DeepMind research scientist; specialized in large-scale robot experimentation, language-conditioned planning, and training infrastructure; leads engineering org building PI's model training systems.
- Lachy GroomCo-Founder & COO2024–presentFormer Stripe product leader and active angel investor (early backer of Figma, Notion, and Ramp); manages fundraising, partnerships, and go-to-market strategy; registered agent on California corporate filings.
- Adnan EsmailCo-Founder2024–presentFormer SVP of Engineering at Anduril Industries and senior Autopilot engineer at Tesla; MIT degrees in Mechanical Engineering, Physics, and EECS; brings deep hardware engineering leadership experience to a software-first company.
- Quan VuongCo-Founder2024–presentFormer Google DeepMind software engineer with expertise in cross-embodiment learning, reinforcement learning, and robotics data pipelines; previously at UC San Diego.
How do you contact Physical Intelligence's leadership?
Physical Intelligence uses the domain physicalintelligence.company for corporate email. The most detected formats are first-name-only (e.g., karol@physicalintelligence.company, ~54%) and first.last (e.g., karol.hausman@physicalintelligence.company, ~40%). For partnership inquiries the company publicly lists partnership@physicalintelligence.company. The addresses below follow verified company format patterns; individual executive addresses have not been separately confirmed as published.
karol@physicalintelligence.companyHow much funding has Physical Intelligence raised?
Physical Intelligence has raised approximately $1.07 billion in confirmed equity funding across three rounds since its March 2024 founding, reaching a $5.6 billion post-money valuation at Series B in November 2025. As of March 2026, the company is reported to be in advanced talks to close a ~$1 billion Series C at an $11 billion valuation.
The seed round closed in March 2024 at $70 million — one of the largest robotics seed rounds on record at the time. Lead investors were Thrive Capital, OpenAI Startup Fund, and Lux Capital, and the round accompanied Physical Intelligence's emergence from stealth. The post-money valuation at seed was approximately $400 million.
The Series A closed in November 2024 at $400 million, pushing the post-money valuation to $2.4 billion. This round came alongside the public debut of the π0 model and was backed by Jeff Bezos personally, OpenAI Startup Fund, Thrive Capital, Lux Capital, Bond Capital, Khosla Ventures, Sequoia Capital, and Redpoint Ventures. It was among the largest single-year robotics fundraises on record and elevated Physical Intelligence to unicorn status fewer than nine months after founding.
The Series B closed in November 2025 at $600 million, led by CapitalG (Alphabet's independent growth fund) at a $5.6 billion post-money valuation. New investors included Index Ventures and T. Rowe Price (a crossover public-market investor making its first Physical Intelligence bet). Jeff Bezos returned, as did Thrive Capital and Lux Capital. Total confirmed equity raised across the three rounds reached approximately $1.07 billion across 14 named institutional and individual backers.
By March 2026, Bloomberg and TechCrunch reported Physical Intelligence in advanced discussions for a further ~$1 billion Series C at an $11+ billion valuation — nearly doubling the company's valuation in under four months. Prospective new investors reported in talks include Founders Fund and Lightspeed Venture Partners, with Thrive Capital and Lux Capital expected to return. That round had not been confirmed as closed as of the publication date of this profile.
How did Physical Intelligence get here?
Physical Intelligence went from research side-project to $5.6 billion company in under two years, driven by landmark model releases, a rapid cadence of open-source contributions, and blockbuster fundraising rounds.
- Q3 2023Research origins — nights-and-weekends collaborationFormer Google DeepMind and UC Berkeley researchers including Hausman, Levine, and Finn begin collaborating on cross-embodiment learning informally, outside their day jobs, building early proofs of concept for a generalist robot policy.
- March 2024Company founded; $70M seed roundPhysical Intelligence incorporated in San Francisco; emerged from stealth with a $70M seed round led by Thrive Capital, OpenAI Startup Fund, and Lux Capital at an ~$400M post-money valuation — one of the largest robotics seed rounds on record. The founding team of seven brings expertise from Google DeepMind, UC Berkeley, Stanford, Anduril Industries, and Stripe.
- October 2024π0 (pi-zero) model unveiledFirst public demo of the π0 Vision-Language-Action generalist policy showing a single model folding laundry, assembling boxes, bussing tables, and performing warehouse induction across multiple robot platforms. Sets a new state-of-the-art on dexterous manipulation benchmarks and validates the cross-embodiment learning thesis.
- November 2024$400M Series A at $2.4B valuationBlockbuster Series A backed by Jeff Bezos personally, OpenAI Startup Fund, Sequoia Capital, Khosla Ventures, Thrive Capital, Lux Capital, Bond Capital, and Redpoint Ventures. Physical Intelligence reaches unicorn status fewer than nine months after founding. One of the largest single-year robotics fundraises on record.
- February 2025π0 open-sourced on GitHub; FAST tokenizer releasedPhysical Intelligence releases π0 model weights and inference code on GitHub (Physical-Intelligence/openpi) and publishes the FAST action tokenization method, enabling approximately 5x faster generalist policy training. Broad academic and research community adoption follows immediately.
- May–September 2025π0.5 released; PyTorch support added to openpiπ0.5 extends π0 with open-world generalization, enabling a mobile manipulator to clean up entirely new kitchens and bedrooms it was never trained on — trained across seven platforms, 68 tasks, and 104 real-world homes. PyTorch support added alongside the existing JAX codebase, opening the model to the broader ML research community.
- November 2025$600M Series B at $5.6B valuation; π*0.6 RL model launchedSeries B led by CapitalG (Alphabet's growth fund) at $5.6B post-money valuation, with Index Ventures and T. Rowe Price joining as new investors. Simultaneously, π*0.6 is released — a VLA model trained via online reinforcement learning that approximately doubles task-completion rates on complex manipulation relative to π0 base.
- April 2026π0.7 steerable compositional model releasedPhysical Intelligence publishes π0.7, a steerable generalist model demonstrating emergent compositional generalization — performing novel multi-step tasks (including using appliances it was never trained on) by recombining learned concepts. Published alongside an arXiv paper (arXiv 2604.15483) and covered by TechCrunch.
- March–June 2026Series C talks: ~$1B at $11B+ valuation reportedBloomberg and TechCrunch report Physical Intelligence in advanced discussions for a ~$1B Series C with Founders Fund and Lightspeed Venture Partners in talks to join alongside returning backers Thrive Capital and Lux Capital, nearly doubling valuation in four months. Round not yet confirmed as closed as of June 2026.
Who are Physical Intelligence's competitors?
Physical Intelligence's competitive set spans hardware-agnostic foundation model labs, vertically integrated humanoid robotics companies, and large technology incumbents building embodied AI on vastly more compute and proprietary data.
- Skild AIPittsburgh-based foundation model rival (raised $1.4B in January 2026 at a $14B valuation — larger than Physical Intelligence's most recent confirmed round); Skild Brain is the most direct head-to-head competitor as a hardware-agnostic generalist policy, backed by SoftBank and NVIDIA Ventures.
- Figure AIVertically integrated humanoid robotics company ($39B valuation as of 2025–2026); builds its own robots AND the AI brain, competing on the full stack rather than licensing intelligence to OEMs — different go-to-market but competes for the same enterprise automation budgets.
- 1X TechnologiesNorwegian-American humanoid robot maker backed by OpenAI targeting home and commercial environments; competes for OEM partnership deals and licensing arrangements that Physical Intelligence's foundation models are also designed to power.
- Covariant (acq. Amazon Robotics, Sep 2024)Acquired by Amazon Robotics in September 2024; its Covariant Brain powers warehouse picking robots with narrower task scope than π0 but is now deeply embedded in Amazon's fulfillment infrastructure — a critical segment Physical Intelligence also targets with partner Ultra.
- Google DeepMind (RT-X / Gemini Robotics)Research organization that originated many of Physical Intelligence's founders and techniques; DeepMind's RT-2 and Gemini Robotics projects are direct academic and commercial rivals with access to far more compute, proprietary data from Alphabet, and no external fundraising constraints.
- Tesla OptimusVertically integrated humanoid program at potentially massive manufacturing scale; Tesla's proprietary data advantage from its vehicle fleet and its gigafactory manufacturing capacity pose a long-term structural threat if Optimus reaches cost parity with specialized industrial robots.
Physical Intelligence — frequently asked questions
