We are thrilled to announce CapitalG’s lead Series B investment in Physical Intelligence. We're honored to partner with Karol Hausman, Lachy Groom, Sergey Levine, Brian Ichter, Chelsea Finn, Adnan Esmail, Quan Vuong, and the rest of the exceptional team at Physical Intelligence as they usher AI into the physical world.

AI’s Physical Frontier

For all the talk of AI automation in the world of software, the reality is that the vast majority of human effort and economic value is spent manipulating atoms, not bits. Performing physical tasks like manufacturing goods, moving materials, preparing food, maintaining infrastructure, and managing households still commands far more of people’s time today than knowledge work or creative pursuits. Beyond the opportunity cost of human potential captured by physical labor, tens of trillions of dollars are spent annually on physical work done by people.

These tasks remain stubbornly human because no machine can match human versatility. The robots of today excel at narrow, repetitive tasks in controlled environments, but when faced with variability, they fail. A warehouse robot that moves boxes can't fold clothing. An assembly line arm built for one product can't switch to the next. Much like Moravec observed in 1988, robots today still stumble dramatically where humans operate with ease. The result is that, for massive industries like manufacturing, logistics, healthcare, food service, agriculture, construction, home services, and so on, the AI revolution to-date has had minimal impact on the bulk of day-to-day operations both at home and at work.

Until recently, the idea of machines genuinely adapting like humans across physical tasks in the real world simply seemed too outrageous to contemplate. But foundation models and the tremendous momentum we have seen in AI over the past 5 years has shattered those assumptions. The same principles and lessons powering the revolution in language, image, and video models are pulling AI into the physical world, unlocking the largest addressable market in history: human physical labor. 

Unlocking this opportunity requires crossing a fundamental divide between where AI has operated to-date – in the ordered realm of bits – to the entropy-laden physical world. Shirts fold differently based on fabric weight and static. Boxes arrive with bent flaps and sticky tape. Coffee pours at varying speeds depending on grind and pressure. Floors have uneven surfaces. Lighting changes throughout the day. No two grasps feel identical. The real world is chaos that refuses to compile, and it punishes systems built on assumptions of consistency.

Learning from Chaos

Physical Intelligence is building the bridge between the worlds of atoms and bits with foundation models that learn to understand and master the chaos of the physical world. Their approach mirrors the revolution across today’s digital-based AI modalities: Gemini does not memorize every sentence; it learns patterns that generalize. Physical Intelligence's models train across diverse robots, tasks, and environments, building intuition about how the physical world works. The central thesis is elegant: Generalization comes from diversity of experience, not perfection on a single task.

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. Manufacturing facilities adapting in days, not quarters. Logistics operations able to handle infinite variety. Homes where robots meaningfully help with daily tasks. Foundation models flip the unit economics entirely: Invest once in the base model, then fine-tune across countless applications with minimal data. 

The opportunity here isn't robotics as traditionally defined. It's every physical task humans currently perform, augmented by intelligent machines. In short, this is one of the largest opportunities for impact and value creation we have ever invested behind at CapitalG.

Physical Intelligence: A Singular Team

Physical Intelligence has assembled what is unquestionably the strongest team in robotics AI. Karol Hausman, Sergey Levine, Brian Ichter, Quan Vuong, and Chelsea Finn are some of the most influential researchers in robotics and reinforcement learning, widely known as the architects of the field's foundational breakthroughs. Lachy Groom is an impactful business and product leader we’ve had the good fortune of observing and supporting as longtime investors in Stripe, where he was an executive for many years. Adnan Esmail is a visionary and high velocity engineering leader with years of experience leading world class hardware engineering teams. They've recruited a team with exceptional talent density: researchers from the world’s top AI labs, engineers who've built production robotics systems, and operators who understand how to scale. In a space with a tremendous amount of noise today, Physical Intelligence is the clear leader and, we believe, the only company with the team, technology, and approach required to build the foundation for physical AI.

The company’s ambition matches the caliber of the team. Physical Intelligence is solving the full stack: model architecture, training recipes combining demonstrations with autonomous learning, data strategy spanning diverse robots and tasks, and infrastructure for reliable deployment. Physical Intelligence is building a robust foundation for real-world demands: control in chaos, recovery from mistakes, wide cross-embodiment, and true generalization.

This is represented in their new π*0.6 model, which represents a field-defining breakthrough. It learns from experience through recap – instruction, coaching, then autonomous practice – handling the compounding errors that plague robotics. When mistakes happen, experts provide real-time corrections while reinforcement learning traces failure back through causality's chain. The results prove it works: π*0.6 operates uninterrupted for hours making espresso drinks, folding novel laundry items in unfamiliar homes, assembling packaging boxes in real factories. Throughput more than doubles. Success rates exceed 90%. This is production-ready reliability, a meaningful leap beyond anything else in robotics today.

Beyond moving quickly, however, the Physical Intelligence team understands the stakes. Unlike many software categories, reliability in the physical world carries mortal stakes. Buildings must stand. Manufacturing lines must run continuously. Physical Intelligence’s approach balances innovation with reliability, pushing boundaries while maintaining the rigor required for real-world deployment. Physical Intelligence is not building a feature or a product. The team is building foundational infrastructure for a new category of intelligence that finally works in the entropy of the real world.

CapitalG’s Investment

We believe Physical Intelligence will be the company that bridges the digital-physical divide, bringing AI's remarkable capabilities to the material world. We're honored to support Karol, Lachy, Sergey, Brian, Chelsea, Adnan, Quan, and the entire team in leading their Series B as they help AI cross the chasm from the digital realm into our physical reality.