Reading time: 26 min
Updated: July 2025
Level: Beginner-Friendly
AI Robotics Company Profile

Physical Intelligence

Building the AI Brain That Makes Any Robot Smarter

Physical Intelligence — known as PI — is the AI research company building foundation models for robots: powerful, general-purpose AI systems that allow robots to learn, adapt, and perform a huge variety of physical tasks. Founded in 2023 by world-leading robotics researchers, PI has raised over $400 million and is redefining what robots can do.

Visit PI Website Wikipedia Wikipedia Meet the Founders
0M+Total Funding
2023Year Founded
0Co-Founders
$0Est. Valuation
About Physical Intelligence

What Is Physical Intelligence?

Imagine you want to teach a robot how to fold laundry. With traditional robotics, you would need a team of engineers to carefully write specific code for every movement — exactly how to pick up a shirt, how to fold each sleeve, where to place it when done. This programming might take months and would only work for that one exact task with that one specific style of shirt. Change the shirt's colour, material, or size, and the whole programme might fail. This is the fundamental limitation that Physical Intelligence was founded to solve — permanently.

Physical Intelligence — commonly known as PI — is an American AI research company founded in 2023. Its core mission is to build foundation models for robots. If that sounds technical, think of it this way: ChatGPT and similar AI systems are "foundation models" for language — they learned from enormous amounts of text and can now answer questions, write essays, translate languages, and do countless other language tasks without being specifically programmed for each one. Physical Intelligence is building the equivalent for the physical world — AI systems that learn from enormous amounts of robot experience and can then apply that learning to perform a wide variety of physical tasks, adapting to new situations without needing to be individually reprogrammed for every scenario.

Simple Analogy: Think of traditional robot programming like teaching someone to ride exactly one specific bicycle model in exactly one specific location. Think of Physical Intelligence's approach like teaching a person to understand the general concept of balance and movement — skills they can then apply to riding any bicycle, a unicycle, a scooter, or even ice skates. The general intelligence transfers to new situations rather than being locked to one specific task.

What makes Physical Intelligence's approach genuinely revolutionary is this: instead of robots needing thousands of hours of custom programming for every new task, their foundation model approach allows a robot to learn a new task from just a small number of demonstrations — or in some cases, no demonstrations at all, by applying knowledge learned from other related tasks. A robot that has learned PI's foundation model can fold towels, pack boxes, make sandwiches, and perform many other manual tasks with far less specific training than conventional approaches would require.

The company was co-founded by four world-class AI and robotics researchers who left positions at Google DeepMind, UC Berkeley, and Stanford to pursue this vision together: Karol Hausman, Sergey Levine, Chelsea Finn, and Brian Ichter. This is one of the most academically distinguished founding teams in the history of AI robotics — each founder is a world-leading expert in their field, and together they represent decades of frontier research in exactly the areas needed to solve this problem. The depth and quality of this founding team is one of the most important reasons Physical Intelligence attracted extraordinary investment almost immediately after launching.

In November 2023, PI launched publicly with a $70 million Series A funding round — a huge amount for a company that had been in existence for less than a year. Just a year later, it raised a further $400 million Series B, valuing the company at approximately $2 billion. The speed and scale of these funding rounds — from an entirely new company with world-renowned academic founders — reflects how excited the technology and investment community is about the potential of foundation models for robotics.

Why It Matters: If Physical Intelligence's research succeeds at scale, it could make robots as broadly capable and adaptable as modern AI language models — able to be deployed in virtually any environment that requires physical manipulation, without months of custom programming for every new task. This could transform manufacturing, healthcare, logistics, agriculture, and ultimately daily life in a way comparable to how the internet transformed communication.

Physical Intelligence's approach is also deliberately different from humanoid robot companies like Figure AI or Tesla Optimus. While those companies focus on building a complete physical robot from scratch, PI focuses on the AI software intelligence — the "brain" — that can be applied to robots of many different types, shapes, and form factors. This means PI's technology can potentially run on a robotic arm in a factory, a mobile robot in a warehouse, a care robot in a hospital, or a humanoid robot in a home — making its approach hardware-agnostic and enormously scalable.

At a Glance

Physical Intelligence — Quick Facts

All the essential information about Physical Intelligence in one clear, easy-to-read place.

Founded
2023
Founders
Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter
Headquarters
San Francisco, California, USA
Industry
AI Robotics / Embodied AI / Foundation Models
Company Type
Private AI Research Company
Research Focus
Foundation Models for Robots & Embodied AI
Total Funding
$400M+ (as of late 2024)
Official Website
pi.website
Wikipedia
Wikipedia
Physical Intelligence on Wikipedia
Main Technology
π0 (Pi-zero) Foundation Model for Robots
Mission
Make robots as generally capable as modern language AI
Status
Active — Rapid Research Progress
The Research Team Behind PI

Co-Founders & Research Leaders

Physical Intelligence was built by four of the world's foremost AI and robotics researchers — here are three of them, whose academic and research contributions form the intellectual foundation of everything PI does.

Karol Hausman, Co-Founder of Physical Intelligence
Wikipedia Wikipedia Profile AI Files Artificial Intelligence Files

Karol Hausman

Co-Founder, Physical Intelligence

Karol Hausman is a robotics and AI researcher who serves as one of the four co-founders of Physical Intelligence. Before PI, he was a Research Scientist at Google DeepMind — one of the world's most respected AI research laboratories — where he worked on some of the most advanced robot learning research conducted anywhere in the world. His work at Google focused on how robots can learn to perform complex, multi-step tasks — moving beyond simple, isolated actions to sequences of actions that require planning and adaptability.

Karol's research has centred on a profound question: how can we build robots that learn general skills rather than narrow, task-specific tricks? This is exactly the question Physical Intelligence was founded to answer, and Karol's years of deep research into robot skill learning, task generalisation, and multi-task robot training directly shaped the approach PI takes to building its foundation models. His expertise in getting robots to learn from data — rather than from explicit programming — is a core pillar of PI's research philosophy.

At Physical Intelligence, Karol brings both the technical depth of his research background and the leadership experience gained working on frontier AI projects at Google. He is one of the principal architects of PI's research direction, helping define what the company is trying to achieve and how it approaches the fundamental scientific challenges of building general-purpose robot intelligence.

Previous RoleResearch Scientist, Google DeepMind
ExpertiseRobot learning, task generalisation, multi-task training
Sergey Levine, Co-Founder of Physical Intelligence
Wikipedia Wikipedia Profile AI Files Artificial Intelligence Files

Sergey Levine

Co-Founder, Physical Intelligence

Professor Sergey Levine is one of the most cited and influential researchers in the fields of machine learning and robotics in the world. He is an Associate Professor at the University of California, Berkeley (UC Berkeley) — one of the world's top computer science universities — where he leads the Robotic AI & Learning (RAIL) Lab, a research group that has produced some of the most important work on robot learning published in the past decade.

Sergey is particularly known for his contributions to deep reinforcement learning — a technique where AI systems learn by trial and experience, similar to how humans learn physical skills through practice — and for his work on offline reinforcement learning, which allows AI to learn from large datasets of previously collected experience rather than only from live interaction. These techniques are fundamental to making robot learning practical at scale, since you cannot always have robots learning in real-time in the real world for every situation they might encounter.

His research at Berkeley has been enormously prolific and impactful. He and his students have published foundational work on learning from demonstrations, meta-learning (where AI learns how to learn faster), and multi-robot systems. Many of the core techniques that Physical Intelligence's π0 foundation model builds on are methods that Sergey helped develop or advance during his academic career. He has published hundreds of papers that have been cited tens of thousands of times — making him one of the most influential individual researchers in AI robotics globally.

InstitutionProfessor, UC Berkeley; Director, RAIL Lab
ExpertiseDeep RL, offline learning, robot skill acquisition
Chelsea Finn, Co-Founder of Physical Intelligence
Wikipedia Wikipedia Profile AI Files Artificial Intelligence Files

Chelsea Finn

Co-Founder, Physical Intelligence

Professor Chelsea Finn is one of the most celebrated young researchers in artificial intelligence and robotics — recognised internationally for her ground-breaking work on meta-learning: teaching AI systems how to learn quickly from very small amounts of data. She is an Associate Professor at Stanford University — one of the world's leading technology universities — and is also affiliated with Google as a Research Scientist.

Chelsea's most famous research contribution is MAML (Model-Agnostic Meta-Learning) — an algorithm she developed during her PhD that allows AI models to learn new tasks extremely quickly from just a few examples, because they have learned a general approach to learning during training. This is enormously significant for robotics: instead of a robot needing hundreds of demonstrations to learn a new task, MAML-inspired approaches can help it learn from just a handful. This fundamental insight — that AI systems can learn to be better learners — is central to Physical Intelligence's approach to making robots rapidly adaptable.

Chelsea has received numerous prestigious awards including the MIT Technology Review 35 Under 35, the NSF CAREER Award, and multiple best paper awards at top AI conferences. She is widely respected not just for the impact of her research but for how she communicates it — making complex ideas accessible to students and the broader scientific community. At Physical Intelligence, her expertise in fast robot learning and generalisation is directly applied to building AI systems that allow robots to acquire new skills rapidly without extensive retraining.

InstitutionAssociate Professor, Stanford University
Key ResearchMAML meta-learning, few-shot robot learning

Fourth Co-Founder: Physical Intelligence was also co-founded by Brian Ichter, another world-class researcher who previously worked at Google and contributed foundational work on robot planning and navigation. Together, the four founders represent an extraordinary concentration of research expertise — arguably the most distinguished founding team in the history of AI robotics.

Company History

Physical Intelligence's Journey

From founding to frontier AI robotics research — here is how PI has grown in an extraordinarily short time.

Early 2023
Company Founded
Karol Hausman, Sergey Levine, Chelsea Finn, and Brian Ichter leave positions at Google DeepMind, UC Berkeley, and Stanford University to co-found Physical Intelligence in San Francisco. The founding team shares a common conviction: the techniques used to build powerful general-purpose language AI — training large foundation models on enormous datasets — can be applied to robotics to create general-purpose robot intelligence. They begin assembling a world-class research team.
November 2023
Public Launch & $70M Series A
Physical Intelligence publicly announces its existence with a $70 million Series A funding round — an extraordinary amount for a company less than a year old. Key investors include Sequoia Capital, Lux Capital, and others who are convinced by both the calibre of the founding team and the soundness of the foundation model approach to robotics. The company publishes early research results demonstrating robots learning new tasks from small amounts of data — evidence that the approach is working as theorised.
2024 — Research
π0 Model Development
Physical Intelligence's research team makes significant progress on its π0 (pi-zero) foundation model — a general-purpose robot intelligence model that can learn to control different types of robots and perform a wide variety of manipulation tasks. The team publishes research showing π0 demonstrating impressive multi-task learning: a single AI model able to fold laundry, stack dishes, box items for shipping, and perform other diverse tasks after seeing demonstrations, without being separately programmed for each task. These results attract enormous attention from the research and robotics communities worldwide.
Late 2024
$400M Series B — $2B Valuation
Physical Intelligence raises a landmark $400 million Series B funding round, valuing the company at approximately $2 billion. The round is led by major technology investors and includes strategic participation from robotics hardware companies interested in PI's software technology. This extraordinary funding for a research company less than two years old reflects investor conviction that PI's approach to foundation models for robots is the right technical direction — and that PI, with its world-class founding team, is the company most likely to make it work at scale.
2024–2025
Expanding Research & Hardware Partnerships
Physical Intelligence expands its research team significantly and begins developing partnerships with robotics hardware companies interested in running PI's foundation models on their robots. The company demonstrates π0 working across multiple different robot hardware platforms — showing the hardware-agnostic nature of its approach. Research advances include improved multi-task performance, faster learning from fewer demonstrations, and better generalisation to tasks and environments not seen during training — key milestones on the path to truly general-purpose robot intelligence.
Future Vision
Towards General Robot Intelligence
PI's roadmap points toward increasingly capable versions of the π0 foundation model, expanded to work with more robot types, more task categories, and with even less training data required for new skills. Long-term, the goal is an AI system for robots equivalent in its generality to what GPT-4 is for language — a single foundation model that enables robots to tackle virtually any physical task in any environment, making AI-powered robotics accessible to industries and applications that cannot currently afford the bespoke programming that traditional robotics requires.
Technology & Products

Physical Intelligence's AI Technology

A detailed look at the technology Physical Intelligence is building and why it represents a genuine breakthrough in robotics.

Embodied AI

Embodied AI means artificial intelligence that exists in a body — an AI system that interacts with the physical world through sensors and actuators (motors and mechanisms that can move and apply force), rather than just processing text or images on a screen. An embodied AI robot can see the world through cameras, feel through force sensors, and affect the world by moving and manipulating objects. Physical Intelligence's research is entirely focused on embodied AI — making AI systems that can sense, reason about, and act in the physical world with the same sophistication that language AI systems bring to text. This is fundamentally different from AI systems like ChatGPT that only process digital information — PI's AI must deal with the messy, unpredictable, physical reality of the real world.

Foundation Models for Robots

A foundation model is a large AI system trained on enormous amounts of data that can then be adapted to many different tasks. GPT-4 is a foundation model for language. DALL-E is a foundation model for images. Physical Intelligence is building the first truly general-purpose foundation model for robot control — a single AI system trained on vast amounts of robot experience data that can then be applied to control many different robots for many different tasks. This is the core of everything PI does. Rather than training a separate AI for each robot task — which is how traditional robotics AI works — PI trains one giant general model and then fine-tunes it for specific applications with much less data. This approach has transformed language AI (one model for everything language-related) and PI believes it will do the same for robotics.

π0 (Pi-Zero) — PI's Core Model

π0 (pronounced "pi-zero") is Physical Intelligence's flagship foundation model for robots — the specific AI system they have built to demonstrate the foundation model approach working for robotics. π0 is a large neural network trained on data from multiple different robots performing multiple different tasks, allowing it to develop general knowledge about how physical manipulation works. Remarkably, a single π0 model can control different types of robotic arms with different configurations, perform diverse tasks including folding clothes, assembling objects, handling packages, and spreading condiments, and learn completely new tasks from just a small number of demonstrations — far fewer than traditional robot learning approaches require. In tests, π0 has demonstrated convincing performance on multiple tasks with a single model, showing that the foundation model approach genuinely works for physical robot control.

Vision Systems

Seeing and understanding the physical world is fundamental to physical intelligence. PI's robots use cameras — typically multiple cameras viewing the same scene from different angles — to perceive their environment. The AI processes these camera images to understand what is in front of it: where objects are in three-dimensional space, what type of object each one is, what state it is in (is the box open or closed? is the cloth folded or crumpled?), and how the current scene relates to the task at hand. PI's vision system uses deep learning — the same technology behind modern image recognition — adapted specifically for understanding scenes in the context of physical manipulation. The vision system is integrated with the control system so that what the robot sees directly informs how it moves, in real time.

Robot Manipulation

Manipulation is the ability to interact with and change the physical world — picking up objects, pressing buttons, turning handles, folding materials, assembling parts. This is one of the hardest problems in robotics because the physical world is enormously varied and unpredictable: different objects have different shapes, weights, textures, and mechanical properties. PI's research focuses specifically on manipulation — teaching robots to use their hands (or grippers) to perform the types of physical interactions that humans do in everyday life and work. The π0 model has demonstrated impressive generalisation in manipulation: trained on many different tasks, it can apply its general understanding of physical manipulation to new objects and new task variations without needing to be specifically programmed for each one.

Multi-Task Learning

Multi-task learning means training a single AI model on multiple different tasks simultaneously, so that knowledge learned on one task can help performance on others. This is the opposite of the traditional robotics approach where a separate model is trained for each specific task. PI's foundation model approach is inherently multi-task: π0 learns from data across many different robot tasks and environments simultaneously, building a rich general understanding of physical manipulation that it can then apply to new tasks. The practical benefit is enormous: instead of needing large amounts of data for every new task a robot needs to learn, the multi-task foundation model can often learn a new task from just a small number of demonstrations, because it already has relevant background knowledge from similar tasks it has seen before.

AI Decision Making

Decision making is what happens between perceiving a situation and acting on it — the reasoning process of figuring out what to do. For a robot, this might involve understanding what the current state of a task is, what the goal is, what sequence of actions is needed to get from here to there, and how to handle unexpected situations that arise along the way. PI's foundation model learns decision-making from experience — by processing enormous amounts of data showing how robots (and humans) make decisions during physical tasks, it develops an implicit understanding of how to plan and execute multi-step physical tasks. This learned decision-making is what allows π0 to handle tasks that require more than just a single action — tasks with multiple steps that need to be executed in the right order, adapting to how the situation changes as each step is completed.

Natural Language Understanding

PI's foundation model integrates language understanding directly with robot control — meaning you can give a robot instructions in plain English and it will understand what you want and act on it. This is possible because PI's model architecture combines vision, language, and action into a single integrated system — following the same approach that has made multimodal language models (like GPT-4V) so powerful. A user might say "fold the towel in the basket and put it on the shelf" and the robot understands this complete instruction — including the sequence of actions required, the relevant objects, and the intended final state — without needing specialised command syntax or precise technical language. Language integration is crucial for making robots practically useful in real environments, where human users expect to communicate naturally.

Learning From Experience

One of PI's core research goals is making robot learning more data-efficient — allowing robots to learn new tasks from fewer demonstrations and less experience than current approaches require. Traditional robot learning often requires hundreds or thousands of examples to train a robot for a single task. PI's approach, enabled by the π0 foundation model, aims to reduce this dramatically: because the foundation model already has rich general knowledge about physical manipulation from pre-training, it needs far less task-specific data to learn a new skill. In research results published by PI, robots have demonstrated learning completely new tasks from just 10-50 demonstrations — a fraction of what traditional approaches need. As the foundation model improves and accumulates more pre-training data, this learning efficiency is expected to improve further.

Hardware Agnostic Platform

A key strategic advantage of PI's approach is that its AI foundation model is not tied to any specific robot hardware. The π0 model has been demonstrated working on multiple different types of robotic arms and end-effectors (the "hands" at the end of robot arms) — showing that the same AI brain can be used across different physical bodies. This hardware-agnostic approach means PI's technology can potentially work with a wide range of existing robot hardware from different manufacturers, rather than requiring companies to adopt a completely new, purpose-built robot. This dramatically increases the addressable market for PI's technology: rather than needing to replace all existing robot infrastructure, companies might be able to upgrade their existing robots by adding PI's AI model — a much more economically attractive proposition for most enterprise customers.

The Process

How Physical Intelligence Works

From a human instruction to a completed physical task — here is every step in PI's AI robot workflow explained in simple language.

1
Human Instruction
The process begins with a human giving the robot an instruction — in plain, natural English. Someone might say "take the shirts from the laundry basket, fold them, and stack them neatly on the shelf," or "pick up the boxes from the conveyor belt and pack them into the large shipping box on the left." Because PI's foundation model integrates language understanding, it can process these spoken or typed instructions as a complete task description — understanding what needs to be done, in what order, with which objects, and to what standard. This natural language instruction capability is what makes PI's robots genuinely useful to non-technical users who should not need to speak in machine language to direct a robot.
2
AI Understanding
The π0 foundation model processes the instruction and builds an internal understanding of the task. This involves identifying the key objects involved (the shirts, the laundry basket, the shelf), understanding the sequence of actions required (pick up, fold, stack), recognising what the successful completion state looks like (neatly stacked shirts on the shelf), and planning how to approach the task given the current state of the environment. This understanding is not based on explicit programming — it emerges from the model's training on enormous amounts of physical task data, through which it has developed a rich implicit understanding of how physical tasks are structured and how to reason about achieving physical goals. The AI then generates a high-level task plan before beginning to execute it.
3
Vision Processing
As the robot begins to engage with the task, its cameras continuously capture images of the environment. The vision system — which is integrated with the foundation model — analyses these images to build a real-time understanding of the scene: where each relevant object is, what its shape and orientation are, what the current state of the task is, and what needs to happen next. For example, when reaching into the laundry basket for a shirt, the vision system needs to identify the shirt among other items, determine its current crumpled state, and figure out how to grasp it appropriately — information that changes with each individual item. The vision processing happens continuously throughout the task, updating the robot's understanding of the scene as things change.
4
Reasoning
Throughout the task, the AI engages in continuous low-level reasoning — making real-time judgments about the current situation and what to do next. This might include: which shirt to pick up first from a pile? How should this particular shirt be grasped given its current position and orientation? Is this step of the task completed correctly? Has something gone wrong (the shirt slipped) and if so, what is the recovery action? This reasoning happens largely implicitly through the neural network processing — the model does not explicitly compute these decisions in a logical, step-by-step way, but rather produces appropriate actions based on what it has learned during training about how to handle similar situations. This implicit reasoning from pattern recognition is one of the key ways deep learning has proven more robust in real-world conditions than explicit logical programming.
5
Action Planning
Based on its vision input and its understanding of the task, the model generates a specific action plan — precise instructions for how the robot's arm and hand should move to achieve the next step. For picking up a shirt, this might specify: move the arm to a position 20cm above the shirt, lower slowly while opening the gripper, close the gripper with this specific force level, lift to this height. These action specifications are generated as a stream of precise numerical commands that the robot's control system translates into actual motor movements. The planning system generates commands continuously — thinking ahead about the next few actions while simultaneously executing the current one — creating smooth, efficient motion that does not stop and start unnecessarily.
6
Robot Movement
The robot's physical systems execute the planned actions — motors receive precise position and force commands, the arm and gripper move with the appropriate speed and force for the task at hand. Force sensors in the gripper provide the AI with real-time feedback about what it is feeling — helping it judge whether it has grasped an object successfully, whether it is applying the right pressure, and whether something unexpected is happening (like an object being heavier than expected). The physical execution of robot movements is a complex control problem in itself, requiring continuous feedback between the planned action, the sensors reporting what is actually happening, and adjustments to compensate for any discrepancy. PI's integrated system handles this closed-loop control seamlessly within the same foundation model framework as the higher-level task planning.
7
Learning From Experience
After completing tasks, the data from those interactions — what the robot saw, what actions it took, and what the outcomes were — can be used to further improve the model. This continuous learning from experience means PI's robots get better over time, not just through new research from the team, but through the accumulation of real-world robot experience data. In the future, a large fleet of robots running PI's foundation model could collectively share their experiences — meaning every robot learns from what all other robots have encountered. This network effect of collective robot learning is one of the most exciting long-term possibilities of the foundation model approach to robotics: as more robots run PI's model and generate more experience data, the intelligence of all robots improves.
Revenue & Strategy

How Physical Intelligence Plans to Generate Revenue

Physical Intelligence is primarily a research company today — but its technology has a clear path to commercial value. Here is the business strategy.

Enterprise Robotics AI

The primary commercial opportunity for Physical Intelligence is selling its AI foundation model — or AI models built on the foundation of π0 — to enterprise robotics customers. A large manufacturer that currently uses robotic arms for specific, pre-programmed tasks could license PI's AI to give those robots dramatically expanded capabilities: the ability to handle a wider variety of parts, adapt to variations in how parts are presented, and learn new tasks with far less programming effort. The economic case is compelling: the cost savings from reducing robot programming time alone — which can run to months and hundreds of thousands of dollars for complex tasks — could justify a significant annual AI software licence fee. PI is well-positioned to target automotive manufacturers, electronics assemblers, consumer goods producers, and other large-scale industrial customers.

AI Research Platform

PI could offer research organisations and robotics companies access to its foundation model infrastructure — cloud-based tools for training robot AI models using PI's pre-trained foundation as a starting point. Similar to how companies like OpenAI offer API access to their language models for developers to build on, PI could offer its robotics AI infrastructure as a platform — allowing other companies and researchers to build AI-powered robot applications without needing to develop the foundational model technology from scratch. This platform model has proven extremely successful in language AI and could work similarly for robotics AI, creating a scalable revenue stream that grows as the platform attracts more users and use cases.

Partnerships & Licensing

Physical Intelligence is actively developing partnerships with robotics hardware companies interested in offering PI's AI capabilities to their customers. A company that manufactures robotic arms, grippers, or mobile robots could license PI's foundation model to offer their hardware with superior AI capabilities compared to competitors — using PI's software differentiation to sell premium hardware. These partnerships create revenue for PI through licensing fees or revenue-sharing arrangements while allowing the hardware partners to compete on AI capability rather than needing to build their own foundation model research from scratch. Hardware-software partnerships of this type have historically been very successful in the technology industry.

Commercial Robotics Solutions

As PI's technology matures, the company may develop end-to-end solutions for specific high-value commercial applications — combining AI model, robot control software, and integration support into a complete package for specific industries. For example, a complete AI-powered laundry folding solution for hotel and healthcare laundry operations, or a complete order fulfilment solution for warehouse operators. These turnkey solutions command higher prices than AI model licences alone and allow PI to capture more of the value its technology creates. The challenge is that going deep in specific verticals requires focused engineering effort, but the payoff is higher revenue per deployment and stronger customer relationships.

Future Revenue Opportunities

The long-term revenue opportunity for Physical Intelligence, if its foundation model approach succeeds, is potentially enormous. Every physical task currently performed by humans in manufacturing, logistics, healthcare, agriculture, retail, and eventually households could theoretically be performed by a robot powered by a general-purpose foundation model like π0. The total labour market for physical tasks globally represents many trillions of dollars of economic activity. If PI's technology achieves even a fraction of that replacement value — and captures a portion of the economic value it creates through licensing — the potential revenue is enormous. This long-term potential is what justifies the extraordinary valuations that venture investors have placed on PI despite its early research stage.

Investment & Growth

Research & Business Growth

Physical Intelligence's funding story is remarkable — from founding to a $2 billion valuation in under two years, driven by the exceptional quality of its founding team and the promise of its technology.

$400M+
Total Funding
Raised across Series A and Series B rounds in under two years — an extraordinary amount for a research-stage company, reflecting enormous investor conviction in the foundation model approach to robotics.
~$2B
Estimated Valuation
The Series B round valued Physical Intelligence at approximately $2 billion — one of the highest valuations ever achieved by an AI robotics company at such an early stage, reflecting the calibre of the team and the scale of the opportunity.
$70M
Series A (Nov 2023)
Raised at launch with Sequoia Capital and Lux Capital as key investors. Remarkable for a company less than a year old — validation of the founding team's exceptional calibre and the promise of the early research results.
$400M
Series B (Late 2024)
A landmark funding round demonstrating continued investor confidence as PI published increasingly impressive research results showing π0 working across diverse robot tasks with genuine multi-task capability.
Sequoia
Key Investor
Sequoia Capital — which backed Apple, Google, Stripe, Airbnb, and ElevenLabs — invested in PI from its Series A, validating PI's potential to become a foundational technology company in the robotics AI market.
π0
Research Milestone
Publication of the π0 foundation model demonstrated that a single general-purpose robot AI model could perform diverse tasks across different robot types — a landmark research result validating the core thesis of Physical Intelligence's approach.
HW Partners
Hardware Ecosystem
Growing partnerships with robotics hardware companies interest in running PI's foundation model on their hardware, establishing PI as a potential AI software layer for the broader robotics hardware ecosystem.

Why Investors Believe: Physical Intelligence's investors are betting on two things simultaneously: the exceptional quality of the founding team (four of the world's foremost AI robotics researchers) and the compelling logic of the foundation model approach (the same technique that transformed language AI should work for robotics AI). When a research company can raise $400M in under two years on the strength of published research results and founding team credentials, it reflects genuine belief that the technology could be transformative.

Real-World Applications

Industries That Physical Intelligence Can Transform

PI's hardware-agnostic foundation model approach means its technology could be applied across virtually every industry that involves physical manipulation tasks.

Manufacturing
Industrial manufacturers can use PI's foundation model to give existing robotic arms dramatically expanded capabilities — handling a wider variety of parts and tolerances, adapting to variations in how components are presented, and learning new assembly tasks with far less programming time and cost. This is valuable for manufacturers who frequently change product lines or who need robots to handle the natural variability of real-world parts.
Warehousing
The diversity of items in a modern fulfilment warehouse — packages come in thousands of different sizes, shapes, and materials — has historically made automation difficult. PI's general-purpose manipulation AI can handle this diversity much better than task-specific robots, potentially enabling robots to pick and pack orders across an entire warehouse product catalogue rather than just a subset of standardised items.
Logistics
Loading, unloading, and sorting in logistics operations require handling enormously diverse packages of different sizes, weights, and fragility. PI's foundation model approach — which handles variation much better than traditional programmed approaches — could make logistics automation more comprehensive, handling tasks that have previously been too variable for robots to manage reliably.
Healthcare
Healthcare environments require careful handling of diverse, often fragile items — medications, medical instruments, patient supplies, and laboratory samples. PI's gentle manipulation capability and ability to work with diverse objects without specific programming for each one makes it potentially valuable for hospital logistics, laboratory automation, and medical device assembly — tasks that currently require human workers because of their diversity and precision requirements.
Agriculture
Agricultural tasks like fruit and vegetable harvesting require delicate manipulation of diverse, naturally variable items in unstructured environments. PI's manipulation AI — trained to handle diverse objects with appropriate force and care — could eventually be applied to agricultural robots that pick produce without damage, adapting to the natural variation in size, shape, and ripeness of real crops in ways that traditional programmed robots cannot.
Retail
Retail environments involve continuous restocking of diverse products in varied packaging, from multiple suppliers with different presentation formats. PI's general manipulation capability could power retail robots that stock shelves, handle returns, organise back-stock, and assist with inventory management across a full product range rather than just specifically engineered product categories.
Hospitality
Hotels and restaurants involve extensive food preparation, serving, and cleaning tasks that require diverse manipulation skills. While current robotics in hospitality is very limited, PI's foundation model approach — capable of generalising from task demonstrations to new task variations — could eventually enable hospitality robots that handle a much wider range of food preparation and service tasks than anything currently possible.
Smart Homes
The home assistant robot market is the ultimate long-term opportunity for general-purpose manipulation AI. A robot that could fold laundry, load and unload the dishwasher, clean surfaces, cook simple meals, and handle the diverse physical tasks of household management would be transformative for families, elderly people, and people with disabilities. PI's foundational research into general-purpose manipulation is precisely the technology that would need to work reliably before such a robot becomes practical.
Space Exploration
Space missions increasingly rely on robots to perform physical tasks in environments too dangerous or distant for human presence. Robots with PI's adaptable, general-purpose manipulation capability could perform maintenance on space stations, set up equipment on lunar and Martian surfaces, and handle the diverse physical tasks of space exploration without needing each specific action to be pre-programmed from Earth — a fundamental limitation of current space robotics.
Scientific Research
Research laboratories need robots that can precisely handle diverse laboratory equipment and materials — pipettes, test tubes, samples, instruments — following diverse experimental protocols. PI's precision manipulation and ability to learn new procedures quickly from demonstrations makes it potentially valuable for laboratory automation, freeing researchers from repetitive manual tasks and enabling experiments to run continuously, even overnight, with robotic assistance.
Competitive Edge

Competitive Advantages

What makes Physical Intelligence stand out in an increasingly competitive field of AI robotics companies.

Foundation Model Approach
PI is one of the few companies applying the foundation model paradigm to physical robot control at a serious research level. If this approach succeeds — as it has for language and image AI — it could create a general-purpose robot intelligence that competitors trained on task-specific approaches cannot easily match.
World-Class Research Team
The founding team — Hausman, Levine, Finn, and Ichter — is arguably the most accomplished group of AI robotics researchers ever assembled at a single company. Their combined research output has defined much of what is known about robot learning. This research depth is difficult to replicate quickly.
Hardware Agnostic
π0 works across different robot types and form factors — it is not tied to one specific robot design. This means PI's technology can potentially run on existing robot hardware across many different companies and industries, dramatically expanding the addressable market compared to companies that build their own proprietary hardware.
Few-Shot Learning
Because the π0 foundation model has rich pre-trained knowledge about physical manipulation, robots running it can learn new tasks from far fewer demonstrations than traditional approaches require. This makes deploying robots for new tasks much faster and cheaper — a key practical advantage for enterprise adoption.
Deep Research Publication
Physical Intelligence publishes its research openly, establishing scientific credibility and attracting top researchers who want to work on frontier problems. This research transparency helps build the reputation needed to attract elite talent and enterprise customers who need confidence in the technical foundations of the AI they adopt.
Data Compounding Effect
As more robots run PI's foundation model and generate experience data, that data can be used to improve the model — creating a data flywheel where more robots lead to better AI, which leads to more robot deployments. This compounding data advantage could become a significant moat as the robot fleet grows.
Strong Investor Backing
With Sequoia Capital and other top-tier investors backing PI with $400M+ at a $2B valuation, the company has significant resources to pursue long-term foundational research without immediate pressure to generate revenue — a crucial advantage for research-intensive deep-tech development.
Software-First Strategy
By focusing on AI software rather than hardware, PI avoids the massive capital expense and operational complexity of manufacturing robots. This asset-light approach allows PI to scale much more efficiently than hardware-focused robotics companies while potentially capturing significant value as the AI intelligence layer across many hardware platforms.
Honest Assessment

Challenges Facing Physical Intelligence

An honest look at the significant challenges PI must navigate on the path to making foundation models for robots a commercial reality.

Robot Safety
A general-purpose robot that can do many different tasks in proximity to humans must be extremely safe — safe enough to work alongside people in factories, hospitals, and eventually homes. Ensuring safety across all the diverse situations a general-purpose robot might encounter is much harder than ensuring safety for a robot doing one specific, pre-defined task. PI's AI must not just perform tasks effectively but must do so without causing harm to people or property in any scenario.
AI Ethics & Accountability
Who is responsible when an AI-powered robot makes a wrong decision that causes harm? Physical Intelligence and its commercial customers will need to navigate evolving legal and regulatory frameworks for AI accountability in physical systems. These frameworks are still being developed in most jurisdictions, and the uncertainty creates legal and reputational risk for early commercial deployments in sensitive environments like healthcare and public spaces.
High Development Cost
Building frontier AI foundation models requires enormous computing resources — the same type of GPU-intensive training that makes large language models expensive to develop. For Physical Intelligence, training a robot foundation model also requires large amounts of robot training data — which is expensive to collect because it requires physical robot hardware running for thousands of hours. Sustaining these research costs while building toward commercial revenue is an ongoing financial challenge.
Hardware Complexity
While PI's software-first approach avoids many hardware challenges, its AI model must still run reliably on physical robot hardware in real environments. Physical hardware fails in ways that software does not — motors wear out, sensors get dirty, cables break. Building AI systems robust enough to perform well despite hardware variability and degradation, across many different robot platforms, is a significant engineering challenge that pure software companies do not usually face.
Competition
Google DeepMind, Meta AI Research, and several well-funded startups are also pursuing foundation models for robotics. Google in particular has enormous resources and a deep history in exactly this research area — many PI founders came from Google. OpenAI is also interested in physical AI. Well-resourced competitors researching the same approaches could potentially match or surpass PI's results with sufficient investment of talent and compute.
Scalability to Production
Demonstrating impressive results in a research environment is different from deploying reliably in a production environment. Real industrial environments have safety requirements, uptime requirements, integration requirements with existing software systems, and operational constraints that research robots do not face. Bridging the gap between impressive research demonstrations and reliable commercial deployment is one of the most persistent challenges in robotics, and it is a challenge PI has not yet fully addressed at scale.
Looking Ahead

The Future of Physical Intelligence

The opportunities in front of Physical Intelligence — if its research succeeds — are potentially transformative for the entire robotics and automation industry.

General Purpose Robots
The ultimate goal: a foundation model so capable that a robot using it can tackle virtually any physical task with minimal task-specific training. Just as GPT-4 can write code, translate languages, and answer medical questions from the same model, PI's long-term vision is a robot AI that can fold clothes, operate machinery, perform surgery preparation, and tend crops — all from one foundational intelligence.
Smart Factories
Future factories powered by PI's AI could have robots that rapidly adapt to changing production requirements — learning new assembly tasks overnight, handling new components without weeks of programming, and continuously improving their performance from accumulated experience. This flexibility could make advanced manufacturing economical at smaller scales than currently possible.
Healthcare Robots
Medical robotics powered by general-purpose manipulation AI could transform hospital operations — handling the enormous volume of physical tasks that currently require human labour, from medication preparation to equipment sterilisation to laboratory sample processing. The precision and adaptability of PI's approach is particularly well-suited to healthcare's requirement for diversity and exactness.
Home Robots
A household robot that can genuinely help with domestic tasks — cooking, cleaning, laundry, errands — requires exactly the type of general-purpose manipulation intelligence that PI is building. The home robotics market could eventually be larger than industrial robotics, and PI's foundation model approach is the most credible path to making household robots practically useful.
Elder Care AI
Ageing populations worldwide face a severe shortage of care workers. Robots capable of assisting elderly people with physical tasks — preparing meals, fetching items, providing reminders, assisting with mobility — could allow many more people to remain independently in their own homes. This deeply human application requires exactly the diverse, gentle manipulation that PI is focused on developing.
AI Workforce
If PI's foundation model approach succeeds at scale, it could provide the AI intelligence layer for an entire generation of commercial robots — the "operating system" of the physical AI economy. Just as Microsoft Windows powered billions of PCs or Android powers billions of smartphones, PI's foundation model could power millions of robots across industries worldwide.
Space & Science
Physical AI with genuine general-purpose capability could transform scientific research and space exploration. Robots that can adaptively execute diverse experiments, maintain equipment, and handle unexpected situations without specific programming for every contingency could extend the reach of science into places and conditions inaccessible to human researchers.
Future AI Economy
The convergence of general-purpose language AI (already transforming knowledge work) and general-purpose physical AI (what PI is building) could unlock unprecedented productivity across the entire economy. Physical Intelligence is positioning itself at the foundation layer of this physical AI economy — if that bet pays off, the company's long-term value could be extraordinary.
Market Landscape

Physical Intelligence vs. Competitors

How does Physical Intelligence's foundation model approach compare to other leading companies in AI robotics?

CompanyFoundedHQApproachFoundation ModelHardware AgnosticEnterpriseKey Strength
Physical Intelligence2023San Francisco 🇺🇸Software AI brain✓ π0 model✓ Multi-robot✓ DevelopingWorld-class research team, foundation model for robots
Figure AI2022Sunnyvale 🇺🇸Full humanoid robot✓ Helix✗ Own robot only✓ BMW deployedOpenAI partnership, commercial deployment
Tesla Optimus2022Austin, TX 🇺🇸Humanoid + AI✓ Dojo AI✗ Optimus only✓ Own factoriesManufacturing scale, compute resources
Boston Dynamics1992Waltham, MA 🇺🇸Physical hardware focus✗ Limited✓ Spot platform✓ IndustrialBest physical robot motion, 30+ years R&D
Agility Robotics2015Salem, OR 🇺🇸Humanoid robot✓ Digit AI✗ Digit only✓ AmazonWarehouse focus, Amazon partnership
Balanced View

Pros & Cons of Physical Intelligence

A fair and honest assessment — what Physical Intelligence does brilliantly and where it still has work to do.

What Physical Intelligence Does Well
  • Most academically distinguished founding team in AI robotics history
  • π0 model demonstrates genuine multi-task robot learning
  • Hardware-agnostic approach scales across many robot types
  • Foundation model approach mirrors the transformation in language AI
  • Few-shot learning from small numbers of demonstrations
  • $400M+ funding from top-tier investors including Sequoia Capital
  • $2B+ valuation validates market confidence in the approach
  • Open research publication builds credibility and talent attraction
  • Software-first strategy avoids expensive hardware manufacturing
Areas of Uncertainty
  • Very early stage — no large-scale commercial deployment yet
  • Primary competitors include Google DeepMind with far greater resources
  • Gap between research results and real-world production deployment remains large
  • Revenue generation is still a future aspiration, not current reality
  • Robot training data is expensive and time-consuming to collect at scale
  • Safety certification for general-purpose robot AI is complex and uncharted
  • Long timelines between research breakthroughs and commercial products
Did You Know?

15 Fascinating Facts About Physical Intelligence

Surprising, remarkable, and inspiring facts about the AI company that may be building the most important technology in robotics.

Fact 01
Physical Intelligence was founded by four people who are each individually considered world-leading researchers in their fields. Having even one scientist of this calibre at a company is unusual; having four as co-founders is arguably unprecedented in the history of AI robotics. Their combined citation count runs into the hundreds of thousands — a measure of how many other researchers have built on their published work.
Fact 02
Physical Intelligence raised $70 million in its Series A round — before the company had even been publicly known to exist for a year. This is extraordinarily rare in technology investment: most companies need years to build credibility before attracting this level of early-stage funding. PI achieved it on the strength of the founding team's reputation and early research results alone.
Fact 03
One of the tasks that PI's π0 model has demonstrated is folding laundry — a task that robotics researchers have been trying to automate reliably for decades. The difficulty of laundry folding is legendary in robotics: fabric is soft, deformable, and behaves unpredictably. The fact that PI's general-purpose model handles it is used as evidence of genuine manipulation capability beyond what specialised approaches have achieved.
Fact 04
Chelsea Finn's MAML (Model-Agnostic Meta-Learning) algorithm — developed during her PhD — is one of the most cited papers in machine learning of the past decade. The core idea: train AI to learn how to learn quickly. This insight, that you can explicitly optimise AI for fast adaptation rather than just final task performance, has influenced the entire field of AI research and is directly applied in PI's approach to fast robot skill acquisition.
Fact 05
The name "Physical Intelligence" is a deliberate contrast to the word "Artificial Intelligence." While AI typically refers to intelligence processing digital information (text, images, data), Physical Intelligence refers to intelligence operating in and through the physical world — embodied AI that can see, touch, and manipulate real objects. The name reflects the company's belief that physical intelligence is the next frontier of AI research.
Fact 06
Physical Intelligence reached a $2 billion valuation before it had even been in existence for two years. For context, it took most legendary technology companies — Google, Facebook, Airbnb — three to eight years to reach this valuation milestone. PI reached it in under two years based on research results and team credentials rather than revenue, reflecting extraordinary investor confidence in the foundation model approach to robotics.
Fact 07
Three of Physical Intelligence's four co-founders left Google DeepMind or Google Research to start the company. Google DeepMind is one of the world's leading AI research organisations — researchers leave well-resourced positions there only for exceptional opportunities. The fact that three world-class researchers were convinced to leave Google together to found PI speaks to their collective conviction in the opportunity and in each other's complementary skills.
Fact 08
The "π0" name for PI's foundational model is a deliberate play on the company name (Physical Intelligence = PI = the Greek letter π), the concept of a "zero-shot" model that can work without task-specific training, and the mathematical constant π (pi) — which is fitting for a company that believes in rigorous mathematical foundations for its AI research. The name is pronounced "pi-zero."
Fact 09
Sergey Levine's RAIL Lab at UC Berkeley has been one of the most productive sources of AI robotics research in the world over the past decade. Multiple techniques that have become standard tools in robot learning research — including offline reinforcement learning, data-driven robot control, and learning from diverse demonstrations — have been developed or significantly advanced by Sergey and his students. Physical Intelligence is, in a sense, the commercialisation of this decade of academic research.
Fact 10
Physical Intelligence's approach is explicitly modelled on the success of large language models. The founders looked at how GPT-4, Claude, and similar systems achieved general language capability through training on enormous diverse datasets, and asked: can we do the same thing for robot control data? The hypothesis is that training on diverse enough robot interaction data will produce the same kind of general capability in the physical domain that LLMs achieved in language.
Fact 11
The π0 model uses a "flow matching" technique for generating robot actions — a mathematical approach from generative AI that has proven effective for generating smooth, precise sequences of robot movements. This import of techniques from image and language generative AI into robot control is characteristic of PI's approach: rather than reinventing robotics from scratch, apply the most advanced AI techniques from adjacent fields to the physical control problem.
Fact 12
Physical Intelligence has demonstrated π0 working on at least seven different robot configurations in published research — different arm designs, different gripper types, different robot base configurations. This multi-robot generalisation is a key proof point for the foundation model approach: the same AI model controlling fundamentally different physical systems successfully, something that would require completely separate programming with traditional approaches.
Fact 13
The laundry folding task that π0 demonstrated required approximately 150 demonstrations to learn — compared to thousands that traditional robot learning approaches would typically require for a task of similar complexity. This data efficiency, made possible by the rich prior knowledge encoded in the foundation model from pre-training, is one of the most practically important properties of PI's approach for commercial deployment.
Fact 14
Physical Intelligence explicitly frames its mission in terms of beneficial impact for humanity — the company's stated goal is to "bring general intelligence into the physical world" for the benefit of people everywhere. This academic, mission-driven culture — rooted in the founders' backgrounds as researchers whose primary motivation was advancing beneficial AI science — distinguishes PI from commercially-driven robotics companies in ways that influence how it conducts research and builds its team.
Fact 15
Physical Intelligence's research is directly relevant to the concept of "embodied AGI" — Artificial General Intelligence that exists in a physical body and can act in the world, not just process digital information. While AGI in language models is still debated, Physical Intelligence is building what could be the first credible path to AGI in the physical domain: a general-purpose intelligence that can tackle virtually any physical task rather than being limited to specific programmed actions.
Common Questions

Frequently Asked Questions

Everything people most commonly want to know about Physical Intelligence — answered simply and clearly.

Physical Intelligence (PI) is an American AI research company founded in 2023 that builds foundation models for robots — powerful general-purpose AI systems that allow robots to learn, adapt, and perform a wide variety of physical tasks without needing to be individually programmed for each one. The company is based in San Francisco and was co-founded by four world-leading AI and robotics researchers: Karol Hausman, Sergey Levine, Chelsea Finn, and Brian Ichter. PI's core product is the π0 (pi-zero) foundation model — a large AI system trained on diverse robot interaction data that can control different types of robots to perform diverse tasks, learning new tasks from relatively small amounts of demonstration data. Think of π0 as the equivalent of GPT-4 for the physical world: just as GPT-4 can perform many different language tasks from a single model, π0 aims to perform many different physical tasks from a single model, using the same general AI intelligence applied to new situations.

Physical Intelligence was co-founded in 2023 by four researchers who represent the highest level of achievement in AI robotics: Karol Hausman, a former research scientist at Google DeepMind focused on robot skill learning and generalisation; Sergey Levine, an Associate Professor at UC Berkeley and one of the world's most cited robotics and machine learning researchers, whose RAIL Lab has produced foundational work in deep reinforcement learning and robot learning; Chelsea Finn, an Associate Professor at Stanford University famous for her MAML meta-learning algorithm that enables AI systems to learn new tasks from just a few examples; and Brian Ichter, another former Google researcher with expertise in robot planning and navigation. All four left prestigious and well-resourced positions at leading research institutions to co-found PI together — a decision that reflects their shared conviction in the foundation model approach to robot intelligence and in each other's complementary expertise.

A foundation model is a large AI system trained on enormous amounts of diverse data that can be applied to many different tasks. In language AI, foundation models like GPT-4 are trained on vast amounts of text and can then perform writing, translation, coding, question answering, and countless other language tasks from the same model, without being specifically programmed for each one. Physical Intelligence is building the equivalent for robots: a foundation model for physical robot control, trained on vast amounts of diverse robot interaction data — video of robots (and humans) performing many different physical tasks — so that it develops a general understanding of physical manipulation that it can apply to new tasks and new situations. The practical benefit is enormous: instead of needing thousands of hours of task-specific programming and data collection for every new robot task, a foundation model approach allows robots to learn new tasks from relatively few demonstrations, because the foundation model already has rich general knowledge about physical manipulation from its training.

Embodied AI means artificial intelligence that exists and acts in a physical body — AI systems that interact with the real physical world through sensors (like cameras that see) and actuators (like motors that move) rather than just processing digital information on a screen. Most AI you are familiar with — like ChatGPT, Midjourney, or recommendation algorithms — is "disembodied" AI: it processes digital information and produces digital outputs, but has no physical presence or ability to interact with the real world. Embodied AI is different: it perceives the physical world (seeing where objects are, feeling how heavy something is), reasons about that perception (deciding what to do with the information), and then acts on the world (moving a robotic arm, walking through a corridor, picking up an object). Physical Intelligence's work is entirely in embodied AI — building AI systems that can sense, reason about, and physically act in the real world. This is considered one of the hardest problems in AI because the physical world is unpredictable, high-stakes, and much more complex to handle than digital data.

π0 (pronounced "pi-zero") is Physical Intelligence's flagship AI foundation model for robot control — the specific system they have built to demonstrate the foundation model approach working for physical robot manipulation. It is a large neural network trained on data from multiple different robots performing multiple different physical tasks, giving it broad general knowledge about how physical manipulation works. π0 can control different types of robotic arms with different configurations, perform diverse tasks including folding clothes, assembling objects, packing boxes, and spreading condiments on food, and learn completely new tasks from relatively small numbers of human demonstrations — far fewer than traditional robot learning approaches typically require. The name is a deliberate combination of the company name (PI), the mathematical constant π, and the concept of "zero-shot" or rapid learning. PI published research on π0 in 2024 showing it performing impressively across multiple tasks simultaneously with a single model, which was considered a significant milestone in AI robotics research.

Physical Intelligence has raised over $400 million in total funding, across two rounds. The first was a $70 million Series A announced in November 2023, when the company publicly launched. The second was a $400 million Series B raised in late 2024 — valuing the company at approximately $2 billion. The Series A investors included Sequoia Capital and Lux Capital among others. The combination of these two rounds, raised in less than two years from a company at the research stage with no commercial revenue, is extraordinary by any standard — reflecting investor conviction in both the calibre of the founding team and the potential scale of the market opportunity for general-purpose robot intelligence. Sequoia Capital, which led or participated in both rounds, is particularly notable — Sequoia has backed some of the most successful technology companies in history, including Apple, Google, Stripe, and Airbnb, and their investment is widely viewed as a strong signal of a company's potential.

Physical Intelligence, Figure AI, and Tesla Optimus are all working in the broad space of AI robotics, but they have fundamentally different approaches. Figure AI and Tesla Optimus are building complete humanoid robots — they design and manufacture the entire physical robot (body, actuators, sensors) and also develop the AI that controls it. Their goal is to sell or deploy a specific physical product: a humanoid robot that works in factories and warehouses. Physical Intelligence, by contrast, is focused almost entirely on the AI software — specifically, the foundation model that serves as the "brain" for robots. PI is not building its own robot body; instead, its π0 model is designed to work across many different robot types made by other companies. PI's bet is that the AI intelligence layer is where the most fundamental value will be created — and that by focusing purely on making the best possible AI brain for robots, without the distraction of also building and manufacturing robot hardware, they can advance the AI capabilities faster and at greater depth than companies trying to solve both problems simultaneously. If PI's approach succeeds, its model could potentially run on Figure AI's robots, or on Boston Dynamics' robots, or on any other robotic platform — positioning PI as the AI software infrastructure for the entire robotics industry.

MAML stands for Model-Agnostic Meta-Learning — an algorithm Chelsea Finn developed during her PhD at UC Berkeley, published in 2017, that has become one of the most influential and widely cited papers in machine learning. The core idea behind MAML is elegant: instead of training an AI model to be good at a specific task, train it to be good at quickly learning any new task from just a few examples. In other words, MAML teaches AI systems how to learn efficiently, rather than just what to learn. The practical result is that a model trained with MAML can adapt to a completely new task using just a handful of examples — sometimes just 5 or 10 demonstrations — because the training process has optimised the model's parameters to be in a state that is easy to fine-tune for new tasks. For robotics, this is enormously valuable: instead of needing thousands of robot demonstrations to teach a robot a new task, MAML-inspired approaches can learn from very few examples. This capability — fast adaptation from minimal data — is directly relevant to PI's goal of making robots that can learn new tasks quickly in commercial deployment environments.

Physical Intelligence's technology has potential applications across every industry that involves physical manipulation tasks performed by workers — which is most of the physical economy. The most immediate beneficiaries are likely to be manufacturing, warehousing, and logistics — industries that already use robots for some tasks but are limited by the inability of current robots to handle diverse, variable tasks without extensive custom programming. PI's general-purpose manipulation AI could make robot automation viable for a much wider range of tasks in these industries than is currently economical. Healthcare is another significant near-term opportunity: hospitals and medical facilities need careful, precise handling of enormously diverse objects and materials, and PI's approach handles diversity much better than task-specific alternatives. Longer term, agriculture, retail, hospitality, and ultimately home assistance are also potential beneficiaries. The home robotics opportunity — a robot that can genuinely help with cooking, cleaning, and household management — requires exactly the type of general-purpose manipulation intelligence that PI is building, and could be the largest single market of all.

Physical Intelligence has a significant historical connection to Google DeepMind — three of its four co-founders came directly from positions at Google DeepMind or Google Research before founding PI. Karol Hausman was a research scientist at Google DeepMind focused on robot learning; Brian Ichter also worked at Google on robot planning and navigation; and Sergey Levine, while primarily based at UC Berkeley, had a deep research collaboration with Google. The founders left Google to build PI because they believed they could advance the foundation model approach to robotics faster and with more focus at a dedicated company than within a large corporation with many competing priorities. Google DeepMind continues to conduct its own research in related areas — including the RT-2 and RT-X robot learning models — meaning PI faces competition from its founders' former employer. However, the founders took their research insights, not Google's code or data, with them, and are pursuing the same scientific direction with an independent team and research agenda. This type of research spin-out — where academic and industry researchers leave to found a company built on insights from their prior work — is common in the technology industry and has produced many successful companies.

Final Thoughts

Conclusion

Physical Intelligence is doing something that might sound simple when stated clearly but is deeply, fundamentally difficult to achieve: building AI that makes robots as adaptable and generally capable as modern language AI has made computers at understanding and generating text. If it succeeds, the impact on the world would be at least as significant — and potentially greater — than what language AI has already achieved.

The company's approach is grounded in a powerful analogy. Before foundation models like GPT-4, language AI was fragmented: you needed separate systems for translation, separate systems for question answering, separate systems for creative writing. Each system was trained for its specific task and could not transfer its knowledge to others. Foundation models changed this completely — one model for everything, learning general language intelligence from enormous diverse data. Physical Intelligence is betting that the same transformation is possible in robotics: instead of separate AI for each robot task, one foundation model that has learned general physical intelligence and can apply it broadly.

The history of AI suggests that whoever builds the foundation model — the general-purpose intelligence layer — for a domain tends to have an enduring advantage. Physical Intelligence is positioning itself to be that foundation for the physical world.

— Summary of Physical Intelligence's core strategic bet

The founding team is extraordinary by any measure. Karol Hausman, Sergey Levine, Chelsea Finn, and Brian Ichter are not aspiring entrepreneurs who wandered into robotics from another field — they are four of the world's foremost researchers in exactly the scientific problems that Physical Intelligence is trying to solve. Sergey Levine's RAIL Lab has produced some of the most important papers in robot learning. Chelsea Finn's MAML algorithm created an entirely new direction in machine learning. These are not impressive credentials for a startup; they are the actual scientific foundations of the field itself. The company is, in a sense, the commercialisation of a decade of frontier academic research by the people who conducted it.

The early results have been encouraging. The π0 foundation model has demonstrated genuine multi-task capability — controlling different robots for different tasks from a single model — and learning new tasks from small numbers of demonstrations. These results are not just marketing demonstrations; they are published research results that the scientific community has scrutinised and found compelling. The fact that Sequoia Capital led both funding rounds — having backed Apple, Google, and Stripe — suggests that investors with high bars and long experience evaluating transformative technology companies see PI as genuinely promising.

For businesses, Physical Intelligence offers a potential future where robot automation becomes dramatically more accessible and flexible. Today, automating a new task with a robot requires months of engineering, large amounts of programming, and significant investment — making it viable only for the highest-volume, most standardised tasks. If PI's foundation model approach works at commercial scale, the same task could potentially be automated in days or weeks from demonstrations, making robot automation economical for a far wider range of manufacturing, logistics, healthcare, and service applications. This democratisation of physical automation would have profound economic implications.

It would be dishonest to suggest there are no challenges. Physical Intelligence is at an early stage, with impressive research results but no large-scale commercial deployment. Building from research prototype to reliable production system is one of the hardest challenges in robotics. The path from "works in a research lab" to "works reliably in a demanding industrial environment" is long and expensive. Well-resourced competitors including Google DeepMind are pursuing similar approaches. And the timelines for transformative robot technology have consistently been longer than optimists have predicted.

But what Physical Intelligence has that most robotics companies do not is intellectual depth at the very frontier of the science. When your co-founders are among the people who invented the techniques the whole field is using, you have a genuine research advantage that money alone cannot easily replicate. And when the approach you are pursuing — foundation models applied to physical AI — is a direct parallel of the most successful paradigm shift in the history of AI, backed by four of the world's best researchers in exactly that area, the probability of success is higher than base rates for robotics startups would suggest.

Physical Intelligence is building something that could matter enormously — not just for businesses looking for automation solutions, but for society as a whole. General-purpose robot intelligence that can be deployed across manufacturing, healthcare, agriculture, and homes could help solve some of the most significant challenges of the coming decades: labour shortages in critical industries, the physical burden of caregiving for ageing populations, the productivity gap between wealthy and developing economies. The potential social impact is enormous. And the team pursuing it is, by any measure, among the most qualified groups of people in the world to have a real chance of achieving it.

Explore Physical Intelligence

Read the research, follow the team, and watch one of the most important AI robotics stories unfold.