Physical AI hits factory floors—but doubts linger

The stack changed when robotics stopped treating perception, reasoning, and control as separate modules and began training one model to close the loop. As a speaker on Cognitive Revolution put it, “the big breakthrough came in 2023 where Google had the … idea of what if we just take a language a VLM and just train it to output robot actions directly,” creating the vision-language-action pattern; IBM Technology said these models add “vision to perceive the world, language to reason about it and action to do something,” while compute efficiency “has gotten ridiculously good,” shrinking 20 million hours of video processing from years to weeks. Even so, the field’s own benchmark was still modest: “RT-2 was a big improvement over previous robotic models, but it was still far less capable than the average human.”

What made that architecture commercially plausible was pairing bigger models with co-designed hardware and richer real-world data. TechCrunch reported Perceptron’s Isaac 0.5 was “fed on a million hours of… general video” and relied heavily on “ego video” plus “UMI video” to connect visual understanding to robotic trajectories, while STMicroelectronics said it “annonce aujourd’hui l’accélération du développement et de l’adoption à l’échelle mondiale de systèmes d’IA physique,” integrating sensors, motor control, depth cameras, Holoscan, and Isaac Sim; Karol Hausman likewise argued, “It became clear that the way to accomplish this is to create an organization whose sole purpose is to solve physical intelligence,” adding that “Over the last two years, the Physical Intelligence (PI) team has been working hard.”

What changed in 2026 was not that robots looked more impressive in demos, but that they started being judged inside operating facilities. Exponential Industry captured the break clearly: “That is a very different picture from 2025. Last year, every robot in the field was remote-controlled, and the fastest finish took 2 hours, 40 minutes and 42 seconds,” whereas in 2026 “autonomy was one of the main things the race was there to test,” marking a shift from supervised validation toward field-ready systems expected to work with less human intervention in real industrial conditions.

That shift showed up in the industrial market itself, not just in research circles. At Hannover Messe 2026, which Exponential Industry described as a “definitive turning point” attended by nearly 100 people across 4,000 exhibitors, major vendors framed physical AI around factory deployment: “NVIDIA and Partners Showcase the Future of AI-Driven Manufacturing,” “Siemens brings AI to the physical world,” and “Rockwell Automation” demonstrated factory system design, while Lenovo explicitly pitched “Production-Scale AI” and claimed “up to 85% faster lead times for manufacturers,” language of operational rollout rather than lab proof.

By late 2026, the money alone made physical AI look less like a speculative lab category and more like an investable market taking shape across multiple layers of industry. The Robot Report, citing The Wall Street Journal, said physical AI firms had already raised more than $23 billion in venture capital in 2026, including Waymo’s $16 billion Series D, Skild AI’s $1.4 billion Series C, NEURA Robotics’ $1.4 billion Series C, and Physical Intelligence’s $1 billion, while investors were explicitly betting that robots could address labor shortages and precision-demanding commercial work.

That capital surge was matched by evidence of platform-scale operating footprints, not just startup decks: Samsara described what its CEO Sanjit Biswas called “the largest AI deployment in the physical world,” with millions of vehicles covering 99% of U.S. roads every day, generating 25 trillion data points a year and helping prevent about 380,000 crashes in the prior year. In the episode, Biswas framed the “AI boom” as “really an infrastructure construction project,” arguing that “the last, call it two decades” had focused on reporting and ingestion, while “in the last two, three years” AIs had become able to act on that data in the physical world. The same period brought fresh financing breadth as investors poured $900 million into XPeng’s humanoid unit, $250 million into Neros, and $200 million each into Gravis Robotics and Gatik.

The pushback is that physical AI still looks more like an early capability curve than a solved industrial platform. TechCrunch captured that skepticism with Harry Mellsop’s line that robotics is in its “GPT 2 era,” adding that “More data and compute will be needed to get over the hump”; in the same vein, critics note that robots may be getting better bodies without the know-how to do value-creating work reliably, and that end-to-end learning still has not consistently produced commercial performance outside narrow, highly structured tasks. That caution persists even alongside bullish market forecasts: Omdia estimates that roughly 60 million industrial robots will be deployed between 2026 and 2035, and ABI Research projects an installed base of 49 million level 3-5 autonomous vehicles (AVs), figures skeptics say do not by themselves resolve questions about robustness, latency, and real-world autonomy.

In industrial settings, the objections become brutally concrete: as NEXT with John Koetsier put it, “If you want to be operating in a factory warehouse, you need to be reliable for three years,” yet “a lot of robots out there are spec'd for payloads of maybe three kilograms,” or “10 kg peak, not 10 kg rated,” and some systems need “on the order of… 10,000 hours of data” for a simple task. Even headline deployments remain bounded: “For ten months, a humanoid robot clocked in at BMW’s Spartanburg, South Carolina plant,” where Figure 02 moved more than 90,000 components and supported over 30,000 BMW X3s—evidence, skeptics say, of narrow success, not broad generalization.