For most of the past decade, the most consequential advances in artificial intelligence have happened in software. Language models, recommendation engines, image classifiers, predictive analytics: the defining products of the AI era have been things you interact with on a screen. The physical world, by contrast, has remained largely resistant to the same transformation. Robots exist, of course, and have for decades. But traditional industrial robots are monuments to rigidity. They’re programmed for a specific task, in a specific location, on a specific component. They excel at repetitive precision. They fail catastrophically when anything changes. Reprogram them and you start the whole process over.
That model is now changing, driven by a convergence that the industry has started calling physical AI: the application of the same foundation-model approaches that transformed language and image generation to the problem of how machines perceive, reason about and act in the physical world. The implications aren’t incremental; they represent a structural shift in what automation can do, where it can operate and who can afford to deploy it.
It’s worth noting that this transformation has been underway longer than most people recognize, in one domain at least. Autonomous vehicles are physical AI in its most mature and commercially validated form. A self-driving car or truck is, in the most literal sense, a robot operating in the real world: perceiving its environment through sensors, reasoning about what it sees, deciding how to act and executing those decisions in real-time, continuously, without human intervention.
The progress here is no longer speculative. Waymo is now delivering over 500,000 paid rides per week across five US cities. Aurora launched fully driverless commercial trucking operations on the Houston-Dallas corridor in 2025 and has since accumulated over 250,000 driverless miles on public highways with zero system-attributed safety incidents, operating routes that a human driver can’t legally complete in a single shift due to hours-of-service regulations. Gatik became the first company in North America to operate fully driverless trucks at a commercial scale for Fortune 50 retailers, including Walmart. These aren’t pilot programs; they’re commercial operations generating real revenue in demanding, uncontrolled environments. Autonomous vehicles have already crossed the threshold that the rest of physical AI is now approaching.
The scale of capital flowing into the broader space reflects how seriously investors are taking that trajectory. Robotics funding reached approximately fourteen billion dollars in 2025, up roughly seventy percent year-over-year, surpassing the previous peak of 2021. That growth isn’t driven by improvements to traditional industrial robots; it’s driven by the emergence of a new category: robots that can learn, generalize and adapt rather than simply execute, in any environment, on any task.
At the foundation of this shift is a problem that has constrained robotics for its entire history: the absence of general intelligence. Traditional robots are programmed, which means every new task requires engineering effort. Every new environment requires re-deployment. Every new object requires re-training. The cost of this brittleness isn’t just financial; it determines the entire market structure. Only large, stable, high-volume operations can justify the investment. Flexibility is simply not available at any price.
What foundation models for robotics offer is the prospect of escaping this constraint. Rather than programming behavior task by task, these models learn generalizable representations of how to interact with the physical world, trained on simulation data, human videos and real-world deployment feedback. The analogy to operating systems isn’t accidental. Just as OSes abstracted hardware so that any software could run on any machine, robotics foundation models are beginning to abstract the hardware so that learned intelligence can run on any body. Autonomous vehicles arrived at this insight first, building AI systems capable of handling the enormous variability of real road conditions rather than programming responses to every possible scenario. The same architectural shift is now propagating across the rest of the physical world.
Skild AI, founded in Pittsburgh in 2023 and now valued at over fourteen billion dollars after a 1.4-billion-dollar funding round, is among the clearest expressions of this logic. The company is building what it calls the Skild Brain, a foundation model designed to be omni-bodied: capable of controlling any robot, regardless of its physical form, without prior knowledge of the specific hardware. Quadrupeds, humanoids, tabletop arms and mobile manipulators can all run the same model. Tested in Pittsburgh’s streets and parks, Skild’s robots navigated environments they’d never seen before, climbed fire escapes and handled manipulation tasks requiring contact reasoning, all without prior mapping or planning. The data flywheel is already turning in production: Skild has deployed on Foxconn’s assembly lines producing Nvidia’s Blackwell GPU servers in Houston, where every task the robots perform feeds back into the model, making it smarter with each deployment. The company grew from zero to approximately thirty million dollars in revenue in a matter of months in 2025, an unusual financial profile for a company still in the foundation-building phase.
The European robotics landscape is developing its own distinctive character, grounded less in platform ambitions and more in deep industrial deployment. Anybotics, founded in 2016 as a spinoff from ETH Zurich’s Robotic Systems Lab and now carrying over 127 million euros in total funding, exemplifies this approach. The company’s Anymal quadruped robots are deployed commercially in some of the most demanding industrial environments on earth: oil and gas installations, chemical plants, nuclear facilities and mining operations.
These aren’t proof-of-concept installations; over two hundred units are in the field, performing thousands of inspections weekly. The robots navigate complex multi-level facilities autonomously, detecting equipment overheating, abnormal vibrations and gas emissions in areas that are either too hazardous or too remote for routine human inspection. In pilot programs with companies including Outokumpu and DSM-Firmenich, Anybotics’ Data Navigator platform analyzed over 2,500 inspections in two weeks. The new Anymal X, the world’s first explosion-certified legged robot designed for use in explosive atmospheres, is coming to market in 2026. This is physical AI solving a concrete operational problem at industrial scale, not a demonstration of what robots might someday do.
A different and equally interesting European bet is being made at the dexterous manipulation layer. Mimic Robotics, another ETH Zurich spinoff, founded in 2024 and backed by Elaia, Speedinvest and the Sequoia Scout Fund, is building AI-driven robotic hands designed to be retrofitted onto standard industrial arms already deployed in factories. The insight driving the company is pointed: Most of the value in physical AI comes not from replacing existing robot arms but from giving them the dexterity to handle the tasks they currently can’t.
An off-the-shelf arm can position accurately; it can’t handle an unfamiliar object, adapt its grip for varying material properties or perform the kind of fine manipulation that still requires human hands in most production lines. Mimic’s foundation model learns from imitation, watching human demonstrations and generalizing from them. The addressable market isn’t the greenfield question of where to deploy robots, but the much larger installed base of industrial systems already in position but still dependent on humans for the tasks that require judgment.
Taken together, these companies illustrate a structural pattern that’s becoming visible across the physical-AI landscape. The layer competition that has defined software platforms for decades is now emerging in robotics. At one layer are the foundation-model companies, building the intelligence that any hardware can run. At another are the hardware platforms, which are increasingly competing on how well they integrate with and benefit from shared intelligence. Between them sits a growing ecosystem of application builders, deploying AI-enabled robots in specific domains and use cases.
What makes this different from previous waves of automation is the nature of the learning loop. Traditional industrial robots don’t improve with use. A physical AI system accumulates real-world deployment data with every task, in every environment, on every hardware variant the model runs on. The more broadly a foundation model is deployed, the better it becomes. The better it becomes, the more attractive it is to deploy. This is the same compounding dynamic that has driven platform advantages in software, now operating in the physical world. It has profound implications for how competitive advantage accrues. The winner in this space won’t necessarily build the best robot; it will build the intelligence layer that the most robots run on, accumulating the broadest and deepest real-world experience data as a consequence.
None of this should obscure the genuine obstacles that remain. The sim-to-real gap, the divergence between how well robots perform in simulation and how they perform in unstructured real environments, is a central unsolved engineering problem. Battery life constraints limit most humanoid robots to between ninety minutes and two hours of operation, far short of what most industrial deployments require. Supply chain concentration is a structural risk: A large proportion of critical components, from actuators to sensors, still originates from a small number of manufacturers in China. And the deployment challenge in uncontrolled environments – hospital wards, construction sites, household settings – remains substantially harder than the warehouse and inspection use cases where physical AI has already demonstrated production-scale viability.
But the trajectory is clear. Physical AI has moved from the research question of whether general robotic intelligence is possible to the engineering question of how to deploy it reliably at scale. That transition from lab to production is where competitive positions are established and the economic value of a technological shift is captured. The companies that prove out real-world deployment now, in industrial inspection, in precision manufacturing, in warehouse logistics, are building the data flywheels, the operational expertise and the customer relationships that will define this market for the decade ahead. To end with Bill Gates: “We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.”
Want to read more like this? Sign up for my newsletter at jan@janbosch.com or follow me on janbosch.com/blog, LinkedIn (linkedin.com/in/janbosch) or X (@JanBosch).
