Physical Intelligence

Physical Intelligence (also known as Pi or π) is an American artificial intelligence robotics company that builds general-purpose foundation models and learning algorithms designed to control any robot for any task. Founded in 2024 and headquartered in San Francisco, California, it is best known for the π (pi) series of vision-language-action models (VLAs), beginning with π₀ in October 2024, which let a single model perceive, reason about, and physically manipulate the world across diverse robot hardware. ^[2][8] By November 2025 the company had raised roughly $1.1 billion across three rounds and was valued at about $5.6 billion. ^[7][8] In March 2026, Bloomberg reported that it was in talks to raise about $1 billion more at a valuation above $11 billion; the company had not announced a completed round as of October 2026. ^[17][18] Its newest published model, π₀.₇, appeared in April 2026, and on October 13, 2026 the company opened a beta of the Pi API, which lets outside developers fine-tune its models and run them through cloud-hosted inference. ^[21][24]

The company was co-founded by a group of robotics and AI researchers and operators: Karol Hausman (CEO, formerly of Google DeepMind), Sergey Levine (UC Berkeley professor), Chelsea Finn (Stanford professor), and Brian Ichter (formerly of Google DeepMind), along with Lachy Groom, Adnan Esmail, and Quan Vuong. ^[9][14] Its stated mission, as described alongside the π₀ release, is to "develop foundation models that can control any robot to perform any task." ^[2]

History

Founding

Physical Intelligence emerged from stealth in March 2024 with a $70 million seed round. ^[9] Its seven co-founders, all named in CapitalG's November 2025 investment note, are Karol Hausman, Sergey Levine, Brian Ichter, Quan Vuong, Chelsea Finn, Lachy Groom, and Adnan Esmail, and all seven appear among the 24 authors of the company's first π₀ paper. ^[1][14] The founding team brought together academic leaders in robot learning, reinforcement learning, and meta-learning with experienced operators from the technology and defense industries.

Several of the founders had worked together on robot learning at Google before the company existed. Hausman, Ichter, Levine, Finn, and Vuong were all co-authors of RT-2 (2023), the Google DeepMind paper that named the category "vision-language-action models" by training a vision-language model on both web data and robot trajectories. ^[34] Groom described the company's origin to TechCrunch in January 2026: he had been following the academic work coming out of Levine's and Finn's labs, heard rumors that they might start something, and tracked down Hausman, who was then a Google DeepMind researcher teaching at Stanford. ^[16] After leaving Stripe, where he was an early employee, Groom had spent about five years as an angel investor before co-founding the company. ^[16]

Karol Hausman, who serves as CEO, previously worked as a Staff Research Scientist at Google DeepMind and held an adjunct professorship at Stanford University, where he co-taught a course on deep reinforcement learning (CS 224R). He earned his PhD from the University of Southern California, where he worked with Gaurav Sukhatme. ^[41] Sergey Levine is an Associate Professor in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, where he leads the Robotic AI and Learning (RAIL) Lab. He received his PhD in computer science from Stanford University in 2014, and his research covers deep reinforcement learning, inverse reinforcement learning, and end-to-end training of neural network policies that combine perception and control. ^[39] Chelsea Finn is an Assistant Professor of Computer Science and Electrical Engineering at Stanford University. ^[40] She is best known for developing Model-Agnostic Meta-Learning (MAML), a technique that trains models to adapt rapidly to new tasks from small amounts of data, published with Pieter Abbeel and Levine in 2017. ^[44] Finn earned her PhD at UC Berkeley under the supervision of Pieter Abbeel and Sergey Levine. ^[43]

Brian Ichter, another co-founder, previously served as a Research Scientist at Google DeepMind and Google Brain. He holds a PhD and MS in Aeronautics and Astronautics from Stanford University and a BS in Aerospace Engineering and BA in Physics from the University of Virginia. ^[38] Adnan Esmail is an MIT graduate who spent four years at Tesla working on technologies for sustainability and self-driving cars before joining Anduril Industries, where a 2020 company profile described him as Anduril's head of electrical. ^[37] CapitalG described him as an engineering leader with years of experience running hardware engineering teams. ^[14] Lachy Groom, a former Stripe executive, contributes business and product leadership. ^[9][14] Quan Vuong came to the company from Google DeepMind, where he contributed to the RT-1 and RT-2 robot models, after doctoral research in deep reinforcement learning, computer vision, and robotics at UC San Diego. ^[16][42] He has described the company's strategy as centered on cross-embodiment learning, so that a new hardware platform does not have to start data collection from scratch. ^[16]

Seed Funding

In March 2024, when it came out of stealth, Physical Intelligence announced a $70 million seed round. The round was led by Thrive Capital, with participation from Khosla Ventures, Lux Capital, OpenAI, and Sequoia Capital. ^[9] CNBC later reported that the seed round "reportedly came in at $70 million with a $400 million valuation." ^[3]

Series A

On November 4, 2024, Physical Intelligence confirmed to CNBC that it had raised $400 million at a post-money valuation of approximately $2.4 billion. ^[3] A company spokesperson named Jeff Bezos, OpenAI, Thrive Capital, and Lux Capital as investors, and Bond Capital said it also invested; CNBC noted that Khosla Ventures and Sequoia Capital were listed as investors on the company's website. ^[3] Crunchbase News, citing The New York Times, reported that the round was led by Jeff Bezos, Lux Capital, and Thrive Capital, with OpenAI, Redpoint Ventures, and Bond also taking part. ^[10] The round put the company's valuation above $1 billion less than a year after it emerged from stealth. ^[3][9]

Series B

In November 2025, the company raised a $600 million Series B round led by CapitalG, Alphabet's growth fund. ^[8][14] The Robot Report said CapitalG led the round "with Lux Capital" and that Bond, Redpoint, and Sequoia Capital also participated, with Jeff Bezos, OpenAI, Redpoint Ventures, T. Rowe Price, and Thrive Capital among previous and returning investors. ^[8] VC News Daily listed Lux Capital, Thrive Capital, Jeff Bezos, Emergence, Index Ventures, and T. Rowe Price as other investors in the round. ^[35] Bloomberg reported that the round valued Physical Intelligence at about $5.6 billion and brought its total funding to roughly $1.1 billion; a briefing from The Information on the same day put the valuation at $5 billion in its headline. ^[7][8][36] In announcing the lead investment, CapitalG partners Jill Greenberg Chase and Manmeet Gujral wrote: "What makes Physical Intelligence's approach transformative is their approach to universally embodied AI: They are building a single generalist intelligence that manifests in any physical form to solve any real-world problem." ^[14]

Funding talks in 2026

On March 27, 2026, Bloomberg reported that Physical Intelligence was in discussions to raise about $1 billion in new funding at a valuation exceeding $11 billion, which would roughly double its valuation from four months earlier. ^[18] According to TechCrunch's account of the Bloomberg report, Founders Fund was set to participate and Lightspeed Venture Partners was in talks to invest alongside returning backers Thrive Capital and Lux Capital; Bloomberg cautioned that the deal was at an early stage and details could change. ^[18] In July 2026, TechCrunch still described the company as having been "reportedly in talks this spring for another $1 billion round at an $11 billion valuation." ^[19] As of October 2026, Physical Intelligence had not announced the round, and its website listed Bond, Jeff Bezos, Khosla Ventures, Lux Capital, OpenAI, Redpoint Ventures, Sequoia Capital, CapitalG, and Thrive Capital as its backers. ^[17]

Anthropic acquisition rumor

In July 2026, a weekend post on X by tech blogger Robert Scoble claimed that Anthropic was acquiring Physical Intelligence, and the rumor spread quickly. ^[19] According to The Information, as reported by TechCrunch, CEO Karol Hausman told employees in a Slack message that the reports were not true; The Information also reported that Anthropic and Physical Intelligence had held acquisition talks in the spring of 2026. ^[19] TechCrunch noted that OpenAI is itself an investor in Physical Intelligence, and that co-founder Lachy Groom did not respond to its request for comment. ^[19] No acquisition was announced, and in October 2026 Physical Intelligence launched its own API beta. ^[21]

Deployment partners and the Pi API

In February 2026 the company published "The Physical Intelligence Layer," arguing that robotics lacks the ready-made intelligence layer that API-based language models give software developers, and presenting results written by two deployment partners. ^[20] Weave, which builds home robots and has deployed laundry-folding robots in Bay Area businesses, reported on a live deployment in a San Francisco laundromat: π₀.₆ increased autonomy compared with π₀.₅, and adding Weave's own data to pre-training reduced missed grasps by 42% and interventions by 50%. ^[20] Ultra, which builds robots for e-commerce order packaging, reported higher success rates with π₀.₆ than with π₀.₅ and showed π₀.₆ packing customer orders for a full shift at 96.4% autonomy. ^[20]

On October 13, 2026, Physical Intelligence announced the beta of the Pi API. Developers upload data, launch fine-tuning on compute and infrastructure that Physical Intelligence manages, and run the fine-tuned models through a hosted inference gateway. ^[21] The company said the latest model available through the API already offered more data-efficient fine-tuning and improved capability for its partners, that it planned to keep providing access to its newest models, and called the launch its "first step in providing the physical intelligence foundation the community can build upon." ^[21] The announcement featured partner work such as Actor's autonomous wheel loader in Utah, which loads material into a dump truck without any operator inside the machine, and early-access users such as Almond, whose Axol dual-arm robot deburred spacers used in its own assembly after the team iterated with the API for four days. ^[21][22] Other early participants included Argon Robotics, Chef Robotics, Deft, Dream Machines, Formic Robotics, Foundry, Frontier Machines, General Instinct, Omakase, and Yondu. ^[22] The company said that most of the applications it showed ran with cloud-hosted inference through the API, and that some deployments ran into network connectivity problems. ^[21] Access to the beta runs through a waitlist. ^[17]

Funding Summary

Technology

Physical Intelligence's technical approach centers on building large-scale foundation models that can serve as general-purpose "brains" for robots. Rather than engineering task-specific control software for each individual robot and scenario, the company trains a single model on diverse robotic data and then fine-tunes it for downstream tasks. This strategy mirrors the approach that large language models (LLMs) use in natural language processing: pre-train on broad data, then specialize. ^[2]

What is a vision-language-action model?

The core of Physical Intelligence's work is the vision-language-action (VLA) model architecture. A VLA model takes in visual observations from cameras, processes natural language instructions, and outputs continuous motor commands that directly control a robot's actuators. This end-to-end approach eliminates the need for separate perception, planning, and control modules, instead learning all three capabilities within a single neural network. ^[1]

Physical Intelligence's VLA models are built on top of pre-trained vision-language models (VLMs), which provide a strong foundation of visual understanding and language comprehension learned from internet-scale image-text data. By extending a VLM with an action generation module, the company's models can follow plain-language instructions (for example, "fold the shirt" or "put the dishes in the sink") and translate them into precise, coordinated motor actions. ^[1]

Flow Matching

A distinctive technical feature of Physical Intelligence's approach is the use of flow matching for action generation. Flow matching is a generative modeling technique related to diffusion models that learns to transform a simple noise distribution into a target distribution of robot actions. Compared to standard autoregressive action prediction, flow matching produces smooth, continuous action trajectories, which is important for dexterous manipulation tasks that require fluid, coordinated movements. ^[1]

In Physical Intelligence's architecture, the flow matching module generates "action chunks," sequences of future actions (typically 50 timesteps at 50 Hz), rather than predicting a single action at a time. This allows the robot to plan short sequences of coordinated movements, resulting in smoother and more reliable execution. ^[1]

Pre-Training and Fine-Tuning

Physical Intelligence follows a two-stage training process. During pre-training, the model is exposed to a large and diverse dataset that includes internet-scale image-text data (through the underlying VLM), open-source robot manipulation datasets such as the Open X-Embodiment dataset, and proprietary datasets collected in-house across multiple robot platforms and dozens of tasks. This broad pre-training gives the model a general understanding of physical manipulation, language instructions, and visual scenes. ^[1]

During fine-tuning, the pre-trained model is adapted to specific tasks or robot configurations using smaller, task-specific datasets. In the π₀ paper, the simplest post-training tasks needed only 5 hours of data, while the most complex used 100 hours or more. ^[1] When it open-sourced π₀, the company said that in its own experiments between 1 and 20 hours of data was sufficient to fine-tune the model to a variety of tasks. ^[4]

Models

π₀ (Pi-Zero)

Physical Intelligence released π₀, its first generalist robot policy, on October 31, 2024. ^[2] The model was described in a research paper titled "π₀: A Vision-Language-Action Flow Model for General Robot Control," authored by a team of 24 researchers including all of the company's co-founders. ^[1] Announcing the model, the company wrote that "we believe this is a first step toward our long-term goal of developing artificial physical intelligence." ^[2]

π₀ is built on top of PaliGemma, a 3-billion-parameter pre-trained vision-language model developed by Google. The model extends PaliGemma with an "action expert" module (adding approximately 300 million parameters, for a total of roughly 3.3 billion parameters) that generates continuous robot actions using flow matching. ^[1] The architecture uses a two-expert design:

• VLM Expert: Initialized from pre-trained PaliGemma weights, this module processes visual observations from cameras and natural language task instructions.
• Action Expert: Randomly initialized and trained from scratch, this module processes proprioceptive state information (such as joint angles) and generates action sequences via flow matching.

A block-wise causal attention mask governs how information flows between these modules: the VLM block attends to itself, the proprioception block attends to itself and the VLM, and the action block attends to all other blocks. ^[1]

π₀ was trained on data from seven robot configurations: single-arm robots (UR5e, Franka), dual-arm configurations (Bimanual UR5e, Bimanual Trossen, Bimanual ARX), and mobile manipulators (Mobile Trossen, Mobile Fibocom). Two of these configuration classes each grouped two similar platforms. ^[1] The training dataset combined proprietary data from over 10,000 hours of robot operation (roughly 903 million timesteps) across 68 tasks with open-source data from the Open X-Embodiment dataset, which contributed about 9.1% of training timesteps. ^[1] Tasks ranged from simple object relocation to complex, multi-step activities such as folding laundry from a hamper, bussing tables (sorting dishes and trash), assembling boxes, bagging groceries, and preparing coffee. ^[1]

In evaluations, π₀ outperformed prior generalist robot models, including OpenVLA and Octo, on five benchmark tasks. The model demonstrated both zero-shot task performance (completing tasks without any task-specific fine-tuning) and strong fine-tuned performance on complex, real-world tasks. ^[1]

Is π₀ open source?

On February 4, 2025, Physical Intelligence open-sourced π₀ by releasing its model weights and code through the openpi repository on GitHub. ^[4][12] The release included weights for both π₀ and the newer π₀-FAST variant, along with example fine-tuning configurations, inference code, and checkpoints for several robot platforms including ALOHA and DROID. ^[4] HuggingFace prepared a PyTorch port of the model alongside the release for developers who prefer PyTorch over the original JAX implementation. ^[4][11] In September 2025 the repository added π₀.₅ checkpoints and native PyTorch support. ^[12] The repository is released under the Apache 2.0 license; as of October 2026 it had about 14,400 stars on GitHub. ^[12]

π₀-FAST

π₀-FAST is an autoregressive variant of π₀ that replaces flow matching with an action tokenization scheme called FAST (Frequency-space Action Sequence Tokenization), introduced in a January 2025 paper. ^[33] Instead of generating continuous action trajectories through iterative denoising, π₀-FAST compresses continuous action sequences into discrete tokens using the Discrete Cosine Transform (DCT). These discrete tokens can then be predicted autoregressively by a standard transformer decoder, similar to how language models generate text. ^[11]

The key advantage of FAST tokenization is training efficiency: π₀-FAST trains approximately five times faster than the diffusion-based π₀ while matching its performance on dexterous and long-horizon manipulation tasks such as laundry folding, table bussing, and grocery bagging. However, one trade-off is that the autoregressive decoding process in π₀-FAST is slower at inference time compared to the flow matching approach used by π₀; the company put its inference cost at about 4 to 5 times higher. ^[4][11]

π₀.₅ (Pi-Zero-Point-Five)

Physical Intelligence published the research paper for π₀.₅ on April 22, 2025 (arXiv:2504.16054). The model represents a significant step toward open-world generalization, meaning the ability to operate effectively in environments that differ substantially from those seen during training. ^[5] The company summarized the goal directly: "We evaluated π₀.₅ by asking it to control mobile manipulators to clean new homes that were never seen in the training data." ^[49]

The central innovation in π₀.₅ is co-training on heterogeneous data sources. While π₀ already combined robot data with VLM pre-training, π₀.₅ goes further by jointly training on data from multiple robots, high-level semantic task predictions, web-scale image and text data, and other sources. This co-training approach teaches the model physical manipulation skills, semantic understanding of objects and environments, the ability to infer task structure, and the ability to transfer behaviors across different robot embodiments. ^[5]

The model maintains a hierarchical design in which the robot is tasked with a general goal and accomplishes it by both predicting semantic subtasks and generating low-level motor actions. Training with web data allows the model to recognize and manipulate objects it has never physically encountered before, because it has learned about those objects from internet images and descriptions. ^[5]

Physical Intelligence ran its final evaluation of π₀.₅ in three real homes that were not part of the training set, where a mobile manipulator completed multi-step household chores such as putting dishes in the sink, placing clothes in a laundry basket, and putting items away in drawers. Controlled comparisons were run in mock home environments. ^[5] The open-source π₀.₅ checkpoints released in openpi in September 2025 were trained with the company's knowledge insulation method. ^[12]

π₀.₆ (Pi-Zero-Point-Six)

π₀.₆ was introduced on November 17, 2025, through a model card published alongside the π*₀.₆ paper. It builds on π₀.₅, keeps its hierarchical design of high-level subtask prediction plus low-level action generation, and changes the pre-trained VLM backbone, the prompt design, and the training data. ^[23] The vision-language backbone is initialized from Gemma 3 4B, the action expert has about 860 million parameters, and the model is trained with knowledge insulation. It can take up to four camera images at 448 by 448 resolution, and it can optionally be prompted with metadata that modulates how a task is performed. ^[23] According to the model card, π₀.₆ produces an action chunk in 63 milliseconds on a single NVIDIA H100 GPU with five denoising steps and three camera inputs, and it performs significantly better than π₀.₅ out of the box, without task-specific fine-tuning. ^[23] Deployment partners Weave and Ultra reported gains from moving to π₀.₆ in early 2026. ^[20]

π*₀.₆ (Pi-Star-Zero-Point-Six)

In November 2025, Physical Intelligence published π*₀.₆, a VLA model that learns from its own deployment experience through reinforcement learning. The paper introduced RECAP (Reinforcement Learning with Experience and Corrections via Advantage-conditioned Policies), a training method that enables VLA models to improve through real-world interaction rather than relying solely on pre-collected demonstration data. ^[6] The company framed the core idea in value-function terms: "actions that result in an increase in the value function, like chess moves that lead to board states from which victory is more likely, are good actions." ^[15]

RECAP incorporates three types of data into the training process: human demonstrations, on-policy data from autonomous robot execution, and expert teleoperated interventions (corrections provided by human operators when the robot makes mistakes during autonomous execution). By conditioning the policy on an advantage signal, the model learns to distinguish between successful and unsuccessful behaviors and gradually improves its performance. ^[6]

π₀.₆ is built on π₀.₆, which the company's blog describes as based on a 5-billion-parameter vision-language model augmented with an action expert. ^[15] In real-world evaluations, the model trained with RECAP more than doubled throughput on some of the most challenging tasks and reduced failure rates by a factor of two or more. On the espresso-making task, throughput more than doubled and the final model reached over a 90% success rate. ^[15] Demonstrated capabilities included folding laundry in real homes, reliably assembling boxes, and preparing espresso drinks using a professional espresso machine. ^[6][15] The company showed π₀.₆ making espresso drinks from 5:30 a.m. to 11:30 p.m., folding 50 different novel laundry items in a new home, and assembling and labeling 59 boxes used for packaging chocolates in a real factory. ^[15]

π₀.₇ (Pi-Zero-Point-Seven)

On April 16, 2026, Physical Intelligence announced π₀.₇, which it described as "a steerable model with emergent capabilities," together with a paper titled "π₀.₇: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities." ^[24][25] The company said the single general-purpose model can perform a range of dexterous tasks with the same performance as fine-tuned specialists, follow new language commands, and perform tasks never seen in its training data, showing what it called the first signs of compositional generalization. ^[24]

The paper describes π₀.₇ as a VLA of about 5 billion parameters: a VLM backbone initialized from the Gemma 3 4B model (including a 400-million-parameter vision encoder), a video history encoder following the company's MEM memory design, and a flow matching action expert with 860 million parameters. ^[25] Its main change is the prompt. During training, π₀.₇ receives multimodal context that describes not only what to do but how to do it: task and subtask language, episode metadata such as speed and quality, control-mode labels (joint or end-effector control), and visual subgoal images. ^[24][25] At run time, the language commands can come from a high-level policy built on the same architecture, and subgoal images come from a lightweight world model initialized from BAGEL, a 14-billion-parameter image understanding, editing, and generation model. ^[25] The company said this labeling lets it train on data that would otherwise hurt performance, such as suboptimal autonomous episodes marked as lower quality or lower speed. ^[24]

In the company's experiments, π₀.₇ made a partial attempt to load a sweet potato into an air fryer from a zero-shot prompt and performed the task much better when coached step by step in language; the coaching data could then be used to fine-tune a high-level policy that ran the task autonomously. ^[24] It also folded laundry on a bimanual UR5e system for which no laundry-folding data had been collected, with a success rate matching the first-attempt success rate of expert teleoperators (who had a mean of 375 hours of teleoperation experience) on that robot. ^[24] By distilling experience gathered during RECAP training into the model with strategy metadata, the single π₀.₇ model matched or exceeded the task-specific π*₀.₆ specialists on laundry folding, espresso making, and box building. ^[24]

Model Summary

Other Research

Besides the numbered models, Physical Intelligence publishes research on methods that feed into them. Members of the founding team appear as authors on most of these papers.

The Robot Olympics post framed the work around Moravec's paradox, the observation that physical skills people find easy remain hard for machines. The company said most of the effort was data collection, under 9 hours for most tasks, and that the two gold-medal tasks it did not solve were physically impossible for its robot, though one could be solved with a small modification. ^[26]

Technical Approach Compared to Other Methods

Physical Intelligence's approach differs from traditional robotics in several ways. Classical robotics systems typically rely on hand-engineered perception pipelines, task-specific planning algorithms, and carefully designed controllers. Each new task or environment generally requires significant engineering effort to adapt. In contrast, Physical Intelligence's learned approach uses a single model that can be adapted to new tasks through data collection and fine-tuning rather than manual engineering.

Compared to other learned approaches in robotics, Physical Intelligence distinguishes itself through its emphasis on generality across robot embodiments. While many prior works in robot learning focus on a single robot platform, Physical Intelligence's models are trained on data from multiple different robots with different morphologies, sensors, and actuators. The use of a pre-trained VLM as the backbone further differentiates the approach from methods that train policies from scratch, since the VLM provides a foundation of visual and linguistic knowledge learned from billions of image-text pairs on the internet. ^[1]

The market research firm Sacra has characterized the company as providing "hardware-agnostic AI models" and likened its approach to an "Android for robots," meaning any robot manufacturer could potentially use Physical Intelligence's models to power its hardware. ^[13] CapitalG described the goal as "a single generalist intelligence that manifests in any physical form." ^[14]

Competition and Industry Context

Physical Intelligence operates in a growing field of companies and research groups working on foundation models for robotics. The competitive landscape includes both well-funded startups and major technology corporations.

Google DeepMind has been a significant player in robotic foundation models, having developed the RT-1, RT-2, and RT-X series of models. RT-2, in particular, demonstrated that a vision-language model could be used to control robots, an approach that influenced subsequent work across the field, including Physical Intelligence's own research. Several Physical Intelligence founders were among RT-2's authors. ^[34] In March 2025, Google DeepMind introduced Gemini Robotics, a model for robotics based on Gemini 2.0. ^[47]

Tesla is developing its Optimus humanoid robot, which leverages the company's extensive experience in data collection and neural network training from its Autopilot and Full Self-Driving programs. Tesla's approach differs in that it is vertically integrated, designing both the robot hardware and the AI software.

Figure AI, another venture-backed startup, focuses on building general-purpose humanoid robots for commercial applications such as manufacturing and logistics. In September 2025, Figure said it had exceeded $1 billion in committed capital in its Series C at a $39 billion post-money valuation. ^[45]

Skild AI, which raised $1.4 billion at a $14 billion valuation in January 2026, is developing a "Skild Brain" foundation model for controlling robots. Like Physical Intelligence, Skild takes a hardware-agnostic approach, but it has already deployed its model commercially and said it generated $30 million in revenue in a few months of 2025. ^[16] Skild has also argued on its blog that most "robotics foundation models" are vision-language models "in disguise" that lack "true physical common sense," because they rely too heavily on internet-scale pretraining rather than physics-based simulation and real robot data. TechCrunch described the two companies' bets as a sharp divide: Skild on a commercial data flywheel, Physical Intelligence on resisting near-term commercialization in favor of more general intelligence. ^[16]

OpenVLA, an open-source VLA model developed by researchers at Stanford, UC Berkeley, and MIT, provides an accessible alternative for research teams looking to build robot policies without commercial software. Physical Intelligence's own open-sourcing of π₀ through the openpi repository places it in direct interaction with the broader open-source robotics AI community. ^[12]

NVIDIA has invested in the physical AI space through its Omniverse simulation platform and Isaac robotics development tools, as well as by releasing its own VLA models. While NVIDIA is not a direct competitor in building robot brain models, its tools and hardware form a critical part of the infrastructure that companies like Physical Intelligence use for training and deployment.

Competitive Landscape

Applications

Physical Intelligence has demonstrated its models on a range of real-world tasks that span household and light-industrial scenarios. Demonstrated applications include:

• Household tasks: Folding laundry (including from hampers), bussing and clearing tables, loading dishwashers, organizing drawers, putting dishes in sinks, and placing clothes in laundry baskets.
• Kitchen tasks: Making coffee, preparing espresso drinks with commercial machines, removing toast from toasters, peeling vegetables, and operating appliances such as an air fryer. ^[15][24]
• Logistics and assembly: Assembling cardboard boxes, bagging groceries, routing cables, placing items in specific locations, and assembling and labeling boxes for packaging chocolates in a factory. ^[15]

The company has said it expects robotic foundation models to enable applications "from homes to hospitals, offices, warehouses, and factories." ^[20] Its own research demonstrations have focused mainly on household and semi-structured environments, while the partner deployments it described in 2026 covered a commercial laundromat (Weave), warehouse order packing (Ultra), and autonomous heavy machinery (Actor). ^[20][21] In January 2026, TechCrunch reported that the company was working with a small number of companies in logistics, grocery, and chocolate making to test whether its systems were good enough for real-world automation, and Quan Vuong said that in some cases they already were. ^[16]

Business Model

Physical Intelligence positions itself as a provider of AI models for robots rather than a robot manufacturer. This software-first, hardware-agnostic strategy means the company does not build or sell robots itself; instead, it develops the "brain" that can be integrated into robots built by other manufacturers. ^[13][14]

In January 2026, co-founder Lachy Groom told TechCrunch that he does not give investors a timeline for commercialization, that the company does not burn much money and spends most of it on compute, and that "there's no limit to how much money we can really put to work." ^[16] The Pi API beta, launched in October 2026, offers hosted fine-tuning and inference of the company's models to outside developers; the announcement did not include pricing. ^[21] TechCrunch described the company in January 2026 as remaining "focused on pure research," and it has not published revenue figures. ^[16]

Team and Organization

Physical Intelligence is headquartered in San Francisco's Mission District. ^[48] When TechCrunch visited in January 2026, the company had about 80 employees and planned to grow, though Groom said he hoped to do so "as slowly as possible." ^[16] Groom said hardware was the hardest part of the work. TechCrunch described a continuous loop in which data is collected on robot stations at the headquarters and at other sites such as warehouses and homes, used to train general-purpose models, and then evaluated back on those stations; the building also contains a test kitchen. ^[16] TechCrunch reported that the robot arms in use at the headquarters sell for about $3,500, a price Levine said includes "an enormous markup" from the vendor. ^[16]

The team includes researchers and engineers from leading AI labs and universities, many of whom have published influential work in deep learning, reinforcement learning, computer vision, and robotic manipulation. The π₀.₇ announcement listed 87 authors. ^[24]

Founders

See also

• List of neolabs
• Neolab
• π0
• π0.5
• π*0.6
• openpi
• Robot foundation model
• Vision-language-action model
• Skild AI
• Config (company)

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al. (2024). "π₀: A Vision-Language-Action Flow Model for General Robot Control." arXiv:2410.24164. arxiv.org/...2410.24164
2. ^1 ^2 ^3 ^4 ^5Physical Intelligence. (2024). "Our First Generalist Policy." Physical Intelligence Blog, October 31, 2024. pi.website/...pi0
3. ^1 ^2 ^3 ^4 ^5 ^6CNBC. (2024). "Jeff Bezos and OpenAI invest in robot startup Physical Intelligence at $2.4 billion valuation." November 4, 2024. cnbc.com/...in-robot-startup-physical-intelligence
4. ^1 ^2 ^3 ^4 ^5Physical Intelligence. (2025). "Open Sourcing π₀." Physical Intelligence Blog, February 4, 2025. pi.website/...openpi
5. ^1 ^2 ^3 ^4Physical Intelligence. (2025). "π₀.₅: A Vision-Language-Action Model with Open-World Generalization." arXiv:2504.16054. arxiv.org/...2504.16054
6. ^1 ^2 ^3Physical Intelligence. (2025). "π*₀.₆: A VLA That Learns From Experience." arXiv:2511.14759. arxiv.org/...2511.14759
7. ^1 ^2Bloomberg. (2025). "Robotics Startup Physical Intelligence Valued at $5.6 Billion in New Funding." November 20, 2025. bloomberg.com/...ued-at-5-6-billion-in-new-funding
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7The Robot Report. (2025). "Physical Intelligence raises $600M to advance robot foundation models." November 2025. therobotreport.com/...ance-robot-foundation-models
9. ^1 ^2 ^3 ^4 ^5 ^6Maginative. (2024). "Physical Intelligence Raises $70M to Build AI-Powered Robots for Any Application." March 2024. maginative.com/...wered-robots-for-any-application
10. ^1 ^2Crunchbase News. (2024). "Robot Brain Startup Physical Intelligence Raises $400M At $2B Valuation." November 2024. news.crunchbase.com/...physical-intelligence-bezos
11. ^1 ^2 ^3HuggingFace. (2025). "π₀ and π₀-FAST: Vision-Language-Action Models for General Robot Control." huggingface.co/...pi0
12. ^1 ^2 ^3 ^4 ^5Physical Intelligence GitHub. "openpi." github.com/...openpi
13. ^1 ^2Sacra. (2025). "Physical Intelligence valuation, funding & news." sacra.com/...physical-intelligence
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9CapitalG. (2025). "Physical Intelligence: Bringing AI Into the Physical World." capitalg.com/...purpose-ai-into-the-physical-world
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7Physical Intelligence. (2025). "A VLA that Learns from Experience." Physical Intelligence Blog, November 2025. pi.website/...pistar06
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14TechCrunch. (2026). "A peek inside Physical Intelligence, the startup building Silicon Valley's buzziest robot brains." January 30, 2026. techcrunch.com/...on-valleys-buzziest-robot-brains
17. ^1 ^2 ^3Physical Intelligence. "Physical Intelligence (π)." Company website (home page, investor list and API waitlist), accessed October 27, 2026.
18. ^1 ^2 ^3 ^4Loizos, C. TechCrunch. (2026). "Physical Intelligence is reportedly in talks to raise $1B, again." March 27, 2026.
19. ^1 ^2 ^3 ^4Loizos, C. TechCrunch. (2026). "The Anthropic-Physical Intelligence rumor roiling AI Twitter." July 21, 2026.
20. ^1 ^2 ^3 ^4 ^5 ^6Physical Intelligence. (2026). "The Physical Intelligence Layer." Physical Intelligence Blog, February 24, 2026.
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Physical Intelligence. (2026). "Frontier Models for Every Robot: Announcing our API beta." Physical Intelligence Blog, October 13, 2026.
22. ^1 ^2The AI Insider. (2026). "Physical Intelligence Opens Pi API Beta for Robot Model Training and Deployment." October 14, 2026.
23. ^1 ^2 ^3Physical Intelligence. (2025). "π0.6 Model Card." November 17, 2025.
24. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Physical Intelligence. (2026). "π0.7: a Steerable Model with Emergent Capabilities." Physical Intelligence Blog, April 16, 2026.
25. ^1 ^2 ^3 ^4Physical Intelligence. (2026). "π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities." Paper (PDF), April 2026.
26. ^1 ^2Physical Intelligence. (2025). "Moravec's Paradox and the Robot Olympics." Physical Intelligence Blog, December 22, 2025.
27. ^Torne, M., Pertsch, K., Walke, H., et al. (2026). "VLAs with Long and Short-Term Memory." Physical Intelligence Research, March 3, 2026.
28. ^Xu, C., Springenberg, J. T., Equi, M., Amin, A., Esmail, A., Levine, S., Ke, L. (2026). "Precise Manipulation with Efficient Online RL." Physical Intelligence Research, March 19, 2026.
29. ^Kareer, S., Pertsch, K., Darpinian, J., Hoffman, J., Xu, D., Levine, S., Finn, C., Nair, S. (2025). "Emergence of Human to Robot Transfer in VLAs." Physical Intelligence Research, December 16, 2025.
30. ^Driess, D., Springenberg, J. T., Ichter, B., et al. (2025). "Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better." arXiv:2505.23705. See also the company's post of May 28, 2025: pi.website/...knowledge_insulation
31. ^Black, K., Galliker, M. Y., Levine, S. (2025). "Real-Time Execution of Action Chunking Flow Policies." arXiv:2506.07339. See also pi.website/...real_time_chunking
32. ^Shi, L. X., Ichter, B., Equi, M., et al. (2025). "Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models." arXiv:2502.19417. See also pi.website/...hirobot
33. ^1 ^2Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., Levine, S. (2025). "FAST: Efficient Action Tokenization for Vision-Language-Action Models." arXiv:2501.09747. See also pi.website/...fast
34. ^1 ^2Brohan, A., et al. (2023). "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." arXiv:2307.15818.
35. ^1 ^2VC News Daily. (2025). "Physical Intelligence Banks $600 Million Funding." November 24, 2025.
36. ^Drew, R. The Information. (2025). "Physical Intelligence Raises $600 Million at $5 Billion Valuation." November 20, 2025 (headline).
37. ^1 ^2Anduril Industries. (2020). "The Heartful Engineer: Adnan Esmail." Anduril Blog, November 17, 2020.
38. ^1 ^2Ichter, B. "Brian Ichter." Personal website.
39. ^1 ^2Levine, S. "Sergey Levine." UC Berkeley EECS faculty page.
40. ^1 ^2Stanford Profiles. "Chelsea Finn." Stanford University.
41. ^1 ^2Hausman, K. "Karol Hausman." Personal website.
42. ^1 ^2AI Engineer. "Quan Vuong." Speaker biography.
43. ^Finn, C. (2018). "Learning to Learn with Gradients." PhD thesis, UC Berkeley, Technical Report UCB/EECS-2018-105.
44. ^1 ^2Finn, C., Abbeel, P., Levine, S. (2017). "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks." arXiv:1703.03400.
45. ^1 ^2Figure. (2025). "Figure Exceeds $1B in Series C Funding at $39B Post-Money Valuation." September 16, 2025.
46. ^1X. (2024). "1X Secures $100M in Series B Funding." 1X Technologies, January 13, 2024.
47. ^Google DeepMind. (2025). "Gemini Robotics brings AI into the physical world." March 12, 2025.
48. ^Physical Intelligence. "Join Us." Company careers page, accessed October 27, 2026.
49. ^Physical Intelligence. (2025). "π0.5: a VLA with Open-World Generalization." Physical Intelligence Blog, April 22, 2025.