Physical Intelligence (π)
Unicorn · $5.6B21 known investors
Physical Intelligence develops general-purpose AI models that enable robots to perform a wide range of manipulation and mobility tasks through foundation models trained on multi-robot, multi-task data. The company creates steerable robotic foundation models designed to generalize across different robot platforms and real-world applications.
Also known as Physical Intelligence · Physical Intelligence (Pi) · Pi · π
Founders & leadership
Investors · 21
.jpeg){kind=logo}
Also in the syndicate · 2
Valuation · disclosed
Disclosed eventsSource: SEC prospectus filings, and round valuations the company or its investors disclosed — follow each entry's link for the claim.
Company profile
researched Aug 2026Physical Intelligence (π) is an artificial intelligence company developing general-purpose foundation models for robots. The company describes itself as a group of engineers, scientists, roboticists and company builders creating learning algorithms for a single model that can control any robot to do any task [0][1][6]. Third-party profiles place its headquarters in San Francisco, California and describe it as building the software or "model layer" for embodied AI rather than robot hardware [5][7].
Its core product line is a series of vision-language-action (VLA) generalist policies. π0, published October 31, 2024, was developed over roughly eight months and combines Internet-scale vision-language pretraining with open-source robot manipulation datasets (Open X-Embodiment) and an in-house dataset of dexterous tasks collected across 8 distinct robot platforms including UR5e, bimanual UR5e, Franka, bimanual Trossen, bimanual ARX, mobile Trossen and mobile Fibocom. It starts from a 3-billion-parameter VLM and adapts it to output low-level motor commands at up to 50 Hz, and can be prompted directly or fine-tuned for specific applications such as laundry folding, coffee making, grocery bagging and table bussing [4]. π0.5, published April 22, 2025, adds co-training on heterogeneous data — multimodal web data, image captioning, visual question answering, object detection, high-level subtask labels and "verbal instruction" demonstrations alongside robot action data — to produce open-world generalization, including controlling a mobile manipulator to clean kitchens and bedrooms in homes not represented in training data [3].
Subsequent releases documented on the company blog include the FAST action tokenizer (January 2025), open-sourced π0 weights and code plus a π0-FAST autoregressive model (February 2025), real-time action chunking for high-latency inference (June 2025), π*0.6 trained with reinforcement learning from experience (November 2025), Multi-Scale Embodied Memory for tasks longer than ten minutes (March 2026), an RL Token method for efficient online reinforcement learning (March 2026), and π0.7, a steerable robotic foundation model (April 2026) [0][2][6]. Reported founding-year and funding details differ between third-party sources: one profile states the company was founded in 2023 with $400M raised at a $2.8B valuation and about 200 employees [5], while another states it was founded in 2024, raised an initial $400M at a $2B valuation and reached $1.1B total funding at a $5.6B valuation after a November 2025 Series B [7].
Founding story
Third-party sources give conflicting accounts. One states the company was founded in 2023 by five researchers in robot learning, naming Sergey Levine (UC Berkeley), Chelsea Finn (Stanford), Karol Hausman (Stanford/Google Brain), Brian Ichter (Google Brain) and Jasmine Hsu (Google Brain) [5]. Another states it was founded in early 2024 by CEO Karol Hausman (former Google DeepMind), Sergey Levine, Chelsea Finn, Lachy Groom (former Stripe), Adnan Esmail, Brian Ichter and Quan Vuong, drawing on researchers from Google DeepMind, Stanford and UC Berkeley [7].
Business model
Physical Intelligence builds no robot hardware; its value proposition is the model layer sitting between natural-language human intent and robot motor commands, intended to be hardware-agnostic across robot embodiments [5]. It develops the software layer enabling diverse robotic platforms to acquire new skills through large-scale multimodal training [7]. One third-party analysis characterizes the company as prioritizing fundamental research over near-term revenue [7].
Traction
The company reports a cadence of model releases from π0 (October 2024) through π0.7 (April 2026), open-sourced π0 weights and code, and a February 2026 post describing partner deployments on real-world problems [0][2]. A third-party profile cites roughly 200 employees and describes π0's dexterity as state of the art on open manipulation and dexterity benchmarks, while noting the seed round reflected the founding team's track record rather than commercial traction [5]. Investors listed by the company include Bond, Jeff Bezos, Khosla Ventures, Lux Capital, OpenAI, Redpoint Ventures, Sequoia Capital, CapitalG and Thrive Capital [0][1][6]; a third-party profile additionally lists Bezos Expeditions, NVentures and T. Rowe Price [7].
Latest developments
The most recent company publication is π0.7, a steerable robotic foundation model with emergent capabilities, dated April 16, 2026 [0][1][2][6]. Preceding it were an online RL method using an extracted "RL Token" (March 19, 2026), Multi-Scale Embodied Memory enabling tasks longer than ten minutes (March 3, 2026) and a post on partner applications titled "The Physical Intelligence Layer" (February 24, 2026) [0][2]. A third-party tracker lists a valuation update dated November 1, 2025 [7].
▸Full profile — market position, technology, go-to-market, geography, history, risks & controversies
Market position
A third-party profile positions Physical Intelligence as a leading proponent of the "one model for all robots" thesis with what it calls the current state-of-the-art VLA model, comparing π0 favorably against OpenVLA, RT-2 and RDT-1B on action generation, cross-embodiment support, language conditioning and dexterity, while noting π0's weights are only partially open [5]. The same source frames the company as a foundational infrastructure bet rather than a near-term revenue play [5].
Rather than training a separate policy per robot and per application, the company trains a single generalist policy on multi-task, multi-robot data so that new tasks require only modest additional data, and inherits semantic and visual understanding from internet-scale vision-language pretraining [4]. π0.5 targets generalization to entirely unseen environments rather than dexterity in known scenes [3]. A third-party profile notes the company builds no hardware, calling the pure-software approach both its strength (hardware-agnostic and scalable) and its risk [5].
Technology
The company's technology centers on vision-language-action models derived from pretrained vision-language models. π0 uses a 3B-parameter VLM backbone adapted for real-time dexterous control, emitting motor commands at up to 50 Hz, and is trained on a cross-embodiment mixture of Open X-Embodiment data, internet-scale pretraining data and proprietary dexterous demonstrations from 8 robot types [4]. π0.5 relies on co-training over heterogeneous data sources and performs both high-level semantic subtask inference (analogous to chain-of-thought) and low-level action prediction through an action expert [3]. Later work adds action tokenization (FAST), real-time action chunking under latency, reinforcement learning from real-world experience (π*0.6, RL Token), long- and short-term memory (Multi-Scale Embodied Memory), and steerability (π0.7) [0][2]. A third-party profile describes π0 as using flow matching for action generation, a ViT-based vision encoder, a language model backbone and an end-to-end trainable flow-matching action decoder [5].
Go-to-market
The company publishes research posts, papers and model releases on its own site and has open-sourced π0 weights and code together with the π0-FAST model, and a February 2026 post describes partners applying its models to real-world problems [0][2]. It lists research@physicalintelligence.company as a contact and maintains an X/Twitter account @physical_int [0][3][4].
The models are intended to control a wide range of robot platforms — single-arm, bimanual and mobile manipulators — for tasks in unstructured environments such as homes, grocery stores, offices and hospitals, in contrast to conventional industrial robots confined to controlled factory and warehouse settings [3][4]. A February 2026 company post refers to partners deploying the models on real-world problems [0][2].
Geography
Third-party profiles list San Francisco, California as the company's headquarters [5][7].
History
π0, described as the company's first generalist policy, was published October 31, 2024 after roughly eight months of development [4]. 2025 brought the FAST tokenizer (January), the open-sourcing of π0 and π0-FAST (February), a step-by-step reasoning method with human-in-the-loop feedback (February), π0.5 with open-world generalization (April), faster-training VLAs (May), real-time action chunking (June), π*0.6 trained with RL (November), work on human-to-robot transfer (December) and fine-tuned models solving "Robot Olympics" manipulation challenges (December) [0][2]. In 2026 the company posted on partner deployments (February), Multi-Scale Embodied Memory (March), efficient online RL via an RL Token (March) and π0.7, a steerable model with emergent capabilities (April) [0][1][2][6]. On funding, one source records a $70M seed in 2023 led by Khosla Ventures and Lux Capital and $400M total at a $2.8B valuation [5]; another records a $400M initial round at a $2B valuation in 2024 led by Jeff Bezos, Lux Capital and Thrive Capital, followed by a $600M Series B led by CapitalG in November 2025 at a $5.6B valuation [7].
Risks & controversies
A third-party analysis notes that the company's pure-software, model-layer approach makes it dependent on the quality and availability of training data it does not collect itself, and that its early valuation reflected team pedigree rather than commercial traction [5]. Sources also conflict on basic facts — founding year (2023 vs 2024), founder list, total funding ($400M vs $1.1B) and valuation ($2.8B vs $5.6B) [5][7]. Source [5] is served from a self-identified staging/QA test environment [5]. The company itself states π0.5 is "far from perfect" and does not always succeed on the first try [3].
Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Timeline · 16
launches, deals, and filingsA steerable robotic foundation model described by the company as exhibiting a step-change in generalization and emergent capabilities.
Method that extracts an 'RL Token' from VLA models to enable fast online reinforcement learning and improve throughput on precise tasks with a few hours of data.
Adds long-term and short-term memory to the company's models, enabling tasks longer than ten minutes.
Company post describing how partners are applying general-purpose physical intelligence models to real-world problems.
Fine-tuned the company's latest model to solve a series of difficult manipulation challenge tasks.
Published findings on how transfer from human videos to robotic tasks emerges in robotic foundation models as they scale.
A method for training generalist policies with reinforcement learning to improve success rate and throughput on real-world tasks.
Series B of $600 million led by Alphabet's CapitalG, bringing total funding past $1 billion at a $5.6 billion valuation.
$600M source ↗
A real-time system for large vision-language-action models that maintains precision and speed under high latency.
A training approach for vision-language-action models aimed at faster training, retention of internet-scale knowledge and better inference properties.
π0.5 extends π0 via co-training on heterogeneous data (multimodal web data, object detection, verbal instructions, subtask commands and robot action data from mobile and static robots), enabling a mobile manipulator to perform tasks such as cleaning kitchens and bedrooms in previously unseen homes.
Published a method enabling robots to reason through complex tasks step by step while incorporating human-in-the-loop feedback.
Released the weights and code for π0 along with the π0-FAST autoregressive model.
A new robot action tokenization method the company says allows training generalist policies 5x faster than previous models.
Physical Intelligence published π0 (pi-zero), a general-purpose robot foundation model developed over roughly eight months that combines Internet-scale vision-language pretraining, open-source manipulation datasets (Open X-Embodiment) and in-house dexterous data from 8 distinct robots, outputting low-level motor commands at up to 50 Hz from a 3B-parameter VLM backbone.
Seed round led by Khosla Ventures and Lux Capital, described by one profile as the largest seed round in robotics at the time of closing.
$70M source ↗
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
▸Research sources · 8
primary sources listed
- Physical Intelligence (π)physicalintelligence.company · web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does Physical Intelligence (π) do?
- Physical Intelligence builds general-purpose vision-language-action foundation models intended to control any robot on any task.
- Who founded Physical Intelligence (π)?
- Physical Intelligence (π) was founded by Karol Hausman, Quan Vuong.
- Who are Physical Intelligence (π)'s investors?
- Physical Intelligence (π)'s investors include Khosla Ventures, CapitalG, Lux Capital, Redpoint Ventures, Sequoia Capital, Thrive Capital, 137 Ventures, Alkeon Innovation and 11 more.