Two weeks ago, I stood in front of a room full of the sharpest minds in manufacturing and supply chain and asked a simple question: “What do you actually mean when you say, ‘physical AI?’”
The term shows up in every pitch deck, every conference keynote, every tech headline. And yet the more popular it becomes, the less precise it gets. So let me offer a working definition—one I think actually holds up.
I define physical AI as how a machine perceives the world around it, how it reasons about what it perceives, and how it acts on that reasoning. Perceive. Reason. Act. Understanding these three pillars cuts through the hype and helps you figure out what’s real.
Reasoning is the pillar furthest along, because the same advances driving language models transfer reasonably well to machines. The real gaps are in perception and action.
Perception is the hardest unsolved problem
At MISUMI’s physical AI summit, Imprint, a striking presentation came from a company tackling fine motor manipulation (getting robot hands to handle objects delicately). The speaker put up a single slide comparing three data sets involving the text for training LLMs, the video for motion capture, and the data for robotic physical manipulation tasks.
While the first two data sets are vast, the third is practically nonexistent by comparison. If you want a robot to pick a strawberry—a task that requires nuanced visual judgment, spatial awareness, and precise motor response—you need robotics training data we simply don’t have yet.
The speaker is betting his company on the belief that closing this data gap is physical AI’s most important task today. I think he’s right. Robotic sensors are nowhere near the capability of human eyes, ears, hands, or proprioception (the body’s sense of where its own limbs are). Perception infrastructure is underrated and underfunded compared to the attention humanoids are getting.
Robots in the human-shaped factory
Most investment is flowing to the third “action” layer, right now. Humanoid robots—general-purpose machines built in roughly human form, meant to work in spaces designed for people— are raising enormous rounds. The demos are impressive. The promise is real.
However, the reality of massive technological advancement is always more complex and difficult than can be conveyed in a demo or a small sample size.
You’ll be disappointed if your plan for deploying humanoids is to put robots in your existing warehouse, factory, or fulfillment center. The amount of perception data alone, the training required for a humanoid to navigate and operate effectively in a space designed for human bodies, is enormous.
Strategic companies are asking: “If I were designing this operation from scratch for a robotic workforce, what would it look like?” This first-principles question requires genuine reimagination, and it’s the only path to the productivity gains justifying the investment.
Robots need their Starlink moment
We’ve been here before. The leased-robot-by-subscription wave of roughly a decade ago had genuinely impressive technology and terrible business models. Companies got stuck in pilot hell, never reaching full deployment. Integration costs ballooned. Interfaces were nightmares. The technology worked in demos and died in deployment. Robotics stayed trapped in R&D.
What breaks that cycle? Look at SpaceX.
Starlink is SpaceX’s revenue engine. By 2025, its satellite internet subscriptions generated $11.4 billion in revenue, accounting for 61% of SpaceX’s total, and it was “the only segment turning a meaningful profit,” according to reporting. The rocket launch business and AI division both posted losses in 2025.
I see how the recurring, high-margin subscription income (once the constellation is in orbit, each new subscriber adds revenue at near-zero marginal cost) funds SpaceX’s costlier, currently unprofitable ventures, including Starship. People paying their monthly internet bill bankroll the audacious long-term vision.
Robots need their Starlink. They need a high-volume, repeatable, revenue-generating use case that can sustain the R&D required to build the next generation of capability. The specific application matters less than the principle: You need a business model that generates real money from real operations, not a perpetual fundraising cycle dependent on private capital.
That’s the arc humanoids need to follow.
What to do right now
If you run a large industrial company, the single most important thing you can do is to get your data ready.
Data is the biggest bottleneck I see, consistently. Get your data house in order for AI adoption, because old systems, siloed data, inconsistent formats, and incomplete records can kill AI initiatives faster than any technology gap.
Next, ask yourself the harder question: Are you trying to automate your current world, or are you willing to reimagine a new one? The companies asking the second question are the ones I’m watching.
Last, before you fund a pilot, ask what its Starlink is: Where is the repeatable revenue that can sustain it past the flashy demo?
At this genuine inflection point, the convergence of reasoning capability, available capital, and physical AI technology is real. I don’t think this is a bubble.
Perceive. Reason. Act. Get those three things right, and build an economic engine underneath them, and the transformation people have been promising for decades might actually arrive.
Dave Evans is president and CEO of MISUMI Americas.