Humanoid robotics is widely framed as a contest of artificial intelligence. McKinsey's analysis suggests it is first an industrial contest. The decisive advantage, however, may belong to organizations that learn to govern something less visible than either machines or models — the authority to decide.
There is a question circulating among regulators in Tokyo, Brussels, and Washington that no performance benchmark can answer. It is not how capable an autonomous system is, or how fast it learns, or how convincingly it handles language. It is simpler and more uncomfortable: when an autonomous system acts, who remains accountable for what it did.
Hold that question for now. It will return.
For the moment, consider the spectacle that dominates the conversation about humanoid robotics. A bipedal machine walks across a stage. It folds laundry, sorts components, navigates a cluttered room. The footage circulates, the valuations climb, and the narrative writes itself: artificial intelligence has finally acquired a body, and the race is on to build the smartest one. The framing feels intuitive. It is also, in the most consequential respects, wrong.
The most clarifying analysis of where humanoid robotics is actually headed came not from a robotics laboratory but from a management consultancy. McKinsey & Company's assessment, presented under the title Humanoid Robots: Crossing the Chasm from Concept to Commercial Reality, begins in an unexpected place. Before discussing robots at all, it discusses the decline of manufacturing. The choice is deliberate, and it reframes the entire question. The humanoid race is not, in the first instance, an AI race. It is an industrial one. And beyond the industrial contest lies a third problem, quieter than both, that will eventually decide which organizations actually benefit from machines that can act on their own.
Why Humanoid Robotics Is Being Misunderstood
The dominant story about humanoid robotics is a story about intelligence. It runs from generative AI, to multimodal models, to the idea of Physical AI — intelligence that leaves the screen and enters the physical world. In that telling, the humanoid robot is the natural endpoint: a general-purpose body for a general-purpose mind. Whoever builds the most capable model, the logic goes, will build the most capable robot, and will win.
This framing is not wrong so much as incomplete, and its incompleteness matters. It concentrates attention on the layer of the problem that is advancing fastest and obscures the layers that are advancing slowest. Vision-language-action models and emerging world models are genuinely impressive. They expand what a robot can perceive, interpret, and attempt. But capability in a demonstration is not the same as capability at scale, and the gap between the two is not primarily a software gap.
Media attention compounds the distortion. A model improvement is easy to show and easy to narrate. A two-minute video of a robot performing a delicate task is more compelling than a spreadsheet describing supplier qualification timelines. The result is a public imagination calibrated to the most photogenic part of the system and blind to the parts that determine whether any of it reaches a factory floor. McKinsey's own framing is blunt on this point: the relevant question is no longer whether these machines can be built. The answer is increasingly yes. The question is whether they can be industrialized — built repeatedly, safely, affordably, and reliably enough to justify the investment.
That shift, from capability to industrialization, is where the real contest begins. And it starts with a problem that has nothing to do with robots.
The Real Driver: Industrial Capacity
For several decades, manufacturing intensity has declined across the Western industrial economies. McKinsey's analysis traces the steady erosion of manufacturing's share of output and employment in the United States, Germany, and other advanced economies, with mainland China absorbing much of the capacity that drained away. This is not a story about robots. It is the economic backdrop against which robots have suddenly become strategically interesting.
The renewed urgency around reshoring — bringing supply chains and production back onshore — is a direct response to that long decline. It is also extraordinarily expensive. McKinsey estimates that achieving the reshoring ambitions of US industry would require on the order of two trillion dollars in capital expenditure. A number of that magnitude does not get spent on nostalgia for an industrial past. It gets spent only if the economics of domestic production can be made to work, and the economics cannot be made to work without automation.
Then there is labor, which is where the boardroom conversation tends to concentrate. The constraint is no longer abstract. In many industrial sectors there are more job openings than applicants. In segments such as warehousing, McKinsey points to annual turnover approaching forty percent — employers recruit, train, and develop workers only to lose them inside a single year, then begin again. Demographic pressure across the advanced economies, including Japan, deepens the shortage rather than easing it. The work increasingly exists; the people to do it do not.
This is the condition that makes Industrial AI and humanoid robotics a strategic priority rather than a curiosity. It is also why the enthusiasm is no longer confined to technologists. McKinsey notes that the number of publicly listed companies naming robotics as central to their strategy has more than doubled since 2023, spanning aerospace and defense, logistics, automotive, consumer manufacturing, and labor services. Automation has migrated from the shop floor to the boardroom, from the operations review to the analyst call.
What is striking, given all this, is McKinsey's assessment that the binding constraint is not robot capability. By its estimate, roughly thirteen percent of current labor hours could be automated with technology that already exists. The obstacle is not what the machines can do. It is everything an organization must change to put them to work.
The Hidden Battlefield: Supply Chains
Here the analysis turns from economics to engineering, and from engineering to geography. McKinsey is explicit that scaling humanoid robotics "is not only an AI problem." Building millions of capable machines requires every component beneath the intelligence — down to the cables — and the supply for some of those components is dangerously thin.
When McKinsey decomposed the bill of materials for a humanoid robot, two subsystems emerged as the critical bottlenecks. The first is sensing: the tactile and force sensors that let a machine feel contact and modulate grip. The second is actuation: the precision drives that move joints with control and repeatability, including harmonic — strain wave — gearing. These are not the glamorous parts of the system. They are also the parts that, historically, have been produced only at research scale, or concentrated among a small number of suppliers in a small number of regions. That concentration is the constraint on industrial scaling.
And the geography is not neutral. Much of a humanoid robot's physical substrate — power electronics, harmonic drives, motors, magnets — overlaps with the value chain of electric vehicles. China, having built the world's leading EV ecosystem under sustained national industrial policy, inherits an enormous head start in robotics from that adjacency. McKinsey observes that the Chinese ecosystem is already moving aggressively to expand capacity precisely at these bottleneck components, which would allow it to hold the control points of the entire supply chain.
This is the part of the story the AI narrative cannot see. The competition for humanoid robotics is, to a significant degree, a competition over harmonic drives and force sensors and the unglamorous capacity to manufacture them at volume. McKinsey frames the decisive capability as "boring industrial muscle" — supplier qualification, supply chain risk management, the patient work of industrialization. For Japan and Korea, whose legacy strengths in mechatronics remain real even after years of relative decline, this is the basis of a credible "China plus one" position: not leadership reclaimed, but indispensability re-established in a world that does not want to depend on a single source.
The point generalizes beyond any one country. Intelligence is becoming abundant and increasingly commoditized. Industrial capacity is scarce, concentrated, and slow to build. In a contest between an abundant input and a scarce one, the scarce one sets the terms.
Why Demos Do Not Win Industries
McKinsey states the conclusion with unusual directness: the humanoid race "is won or lost not by demos. It's won or lost by the capacity to make real industrial products come to life."
The distance between a successful prototype and a commercial deployment is wider than it appears, and it is measured in dimensions that rarely feature in promotional footage. A robot that performs brilliantly for a few hours is an engineering achievement and a commercial liability. McKinsey identifies sustained uptime — the ability to perform reliably across an entire shift, day after day — as one of the bridges the industry must still cross. Safety is another. Industrial robots and collaborative robots operate within mature standards developed over years; for humanoids, comparable standards are still being written, and McKinsey notes that such frameworks have historically taken more than seven years to develop and adopt. The industry is now trying to compress that timeline to a couple of years, which is itself a tacit admission of how much rests on it.
Cost is the third bridge, and the one a chief financial officer ultimately has to sign. McKinsey draws the analogy to batteries in electric vehicles: adoption at scale depends on a steep, sustained reduction in unit cost, achieved through industrializing the supply chain, modularizing the design, and stripping out part complexity. Capability is the fourth — greater dexterity and autonomous mobility, because the tasks that remain unautomated tend to be precisely the ones that involve assembly, deformable materials, and objects that tangle.
None of these four bridges is primarily a question of model intelligence. They are questions of engineering discipline, manufacturing scale, regulatory maturity, and capital. This is why the company with the most impressive demonstration is not necessarily the company that will dominate the industry. Generational technologies — railroads, electrification, the internet — were not won by the first impressive demonstration. They were won by whoever could industrialize, distribute, and operate at scale. McKinsey expects general-purpose robotics to follow the same pattern, and projects that the general robotics market, including humanoids, could reach roughly 370 billion dollars by the end of the next decade — a figure many times larger than today's industrial-arm market, yet smaller than the trillion-dollar narratives suggest, because it counts what is real rather than what is hoped for.
So the first reframing is established. The humanoid race is an industrial race, not an AI race. But there is a second reframing, and it is the one almost no one is discussing.
The Governance Question Nobody Is Asking
Recall the question from the opening — the one regulators keep asking that no benchmark can answer. It is time to take it up.
While the industry concentrates on supply chains and uptime, governments have been quietly converging on a different concern. Across jurisdictions, regulators are beginning to require that autonomous AI systems remain subject to meaningful human oversight — not as a courtesy to public anxiety, but as a structural safeguard against malfunction, opacity, and the diffusion of responsibility. The emphasis is consistent: a human must remain genuinely able to supervise and, where necessary, intervene.
Japan's framework is instructive precisely because it is measured rather than alarmist. The AI Guidelines for Business (Version 1.2), issued jointly by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry, sets out the importance of appropriate human oversight and intervention in the operation of AI systems. The language is not prohibitive. It does not forbid autonomy. It insists that autonomy be accompanied by a structure in which human judgment retains a defined place.
Read in the context of a factory full of humanoid robots, that requirement stops being a compliance footnote and becomes a design problem. If the machines on the floor are sensors and actuators and drives, why does a governance instrument reach past all of that hardware to specify the role of human judgment? The answer is the hinge on which this entire analysis turns. Regulators have noticed something the industrial conversation has not yet fully absorbed: as decisions migrate from people to autonomous systems, the question of who is accountable for those decisions does not migrate with them. It is left behind, unowned.
The problem, in other words, is no longer simply technological. It is institutional. And it sharpens as the machines multiply.
When Factories Become Judgment Networks
Picture a single humanoid robot on a production line and ask a deceptively simple question: on whose instruction is it acting?
The honest answer is that no single instruction exists. The robot's behavior emerges from a lattice of interacting systems. An enterprise resource planning (ERP) system sets the production plan. A manufacturing execution system (MES) translates that plan into operations on the floor. A supply chain management (SCM) system governs what arrives and when. Increasingly, AI Agents sit above these systems, interpreting conditions and issuing instructions autonomously rather than waiting for a human to act. And around all of it, people remain: the operator at the station, the floor supervisor watching the line, the manager who answers for the outcome.
Each of these participants — human and machine — exercises judgment. The robot's motion is the visible output of an invisible network of decisions distributed across software and people. Under normal conditions, no one notices the network at all. It surfaces only when something goes wrong. When a humanoid robot mishandles a component, damages a product, or behaves outside its intended envelope, the organization is forced to ask a question it never thought to ask in advance: who decided?
That question is harder to answer than it should be, because the future factory is not a collection of machines. It is a network of distributed judgment, in which humans, AI Agents, humanoid robots, and enterprise systems are continuously connected, one decision feeding the next. McKinsey's concerns about uptime and safety take on a different meaning at this scale. The risk is not only that one machine fails. It is that an error in judgment propagates through the network — amplified, repeated, and untraceable — before any human recognizes it. Beneath the physical wiring diagram of such a factory runs a second diagram that no one has drawn: the wiring of judgment itself. Add machines to an ungoverned version of that network and it does not become more intelligent. It becomes less controllable.
This is the governance challenge that Industrial AI and Physical AI create and that conventional AI Governance, focused on models and data, does not fully address. It is not a question about how the models behave. It is a question about who, in a system of distributed decision-making, holds authority — and who answers for its consequences.
Beyond Human-in-the-Loop
The reflexive response to all of this is to invoke Human-in-the-Loop (HITL). Keep a person in the decision path. Require human oversight. Demand human approval at critical junctures. Surely, the reasoning goes, a human in the circuit keeps the system safe.
The instinct is sound. The implementation is where it fails. Consider the recurring concern from the opening, and the human oversight that Japan's AI Guidelines for Business (Version 1.2) and comparable frameworks elsewhere call for. The requirement is rarely the problem. The form it takes in practice is.
What organizations frequently build is approval that is human in name only. A screen presents a conclusion. A button offers to approve it. A person clicks. But the conclusion was assembled by an AI system from data and reasoning the human neither produced nor fully examined, often under time pressure that makes genuine scrutiny impractical. The human is in the loop, but the loop runs through them rather than depending on them. Oversight becomes ratification.
This produces a quiet but serious distortion, and it is the heart of the matter. Authority and responsibility come apart. Effective authority — the capacity to determine the outcome — migrates to the AI system that shapes the recommendation. Responsibility — the obligation to answer for the outcome — remains with the human who approved it. When something goes wrong, a person is accountable for a decision they did not meaningfully make. A governance framework that mandates human involvement without structuring it does not resolve this divergence. It can entrench it, converting the human from a safeguard into an absorber of liability for choices made elsewhere.
The lesson is that keeping a human in the loop is not, by itself, governance. The decisive question is not whether a human is present in the decision path, but where the boundary between human and machine authority is deliberately drawn — and whether responsibility has been placed on the same side of that boundary as the authority it is meant to accompany. Naming that boundary, and designing it, is a different kind of work than adding an approval step. It is the work the rest of this analysis has been moving toward.
Decision Boundaries and Judgment Governance
The concept that addresses this divergence directly is the Decision Boundary. It treats the allocation of judgment between humans and machines not as something that emerges informally from workflow, but as something to be specified by design. A Decision Boundary makes explicit what distributed systems tend to leave implicit: what class of judgment is being delegated, who is the decision-maker, how much authority is delegated to an AI Agent or a robot, where that delegation stops, who bears responsibility for the result, how a situation that exceeds the system's envelope is escalated, and under what conditions a human may override the machine's determination.
It is essential to be precise about what such a boundary is. Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority. A safety interlock that halts a robot when a sensor trips is an operational threshold. A Decision Boundary is something prior and more durable: a statement of who, within the institution, is entitled to decide a given class of question, and who must answer for it. The first is engineering. The second is governance.
Designing and maintaining these boundaries across a network of distributed judgment is the discipline this analysis has been circling. Conventional AI Governance — concerned with the fairness, transparency, and data practices of models — remains necessary, but it is not sufficient for a world in which Physical AI acts on the factory floor. What that world additionally requires is the governance of judgment as such: ensuring that every consequential decision has an identifiable owner, that the chain of decisions can be reconstructed after the fact, and that authority and accountability remain joined rather than drifting apart. This challenge sits adjacent to AI Governance but cannot be fully solved by model governance alone. This is Judgment Governance — distinct from model governance, overlapping with it, and not reducible to it.
Reconstruction depends on record. This is where decision logs earn their place. Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes. A log of what a system produced is an operational artifact. A record of who held authority, what was delegated, where the boundary sat, and how an outcome was reached is a governance artifact — the connective tissue that allows responsibility to survive the passage of a decision through many hands, human and machine.
These concepts — the Decision Boundary, Judgment Governance, the decision log — are facets of a single underlying idea, which is that judgment itself can be treated as an object of design. Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate. That distinction is the whole argument in miniature. The aim is not a smarter recommendation. It is a structure in which the question who decided always has an answer, and in which that answer carries the weight of legitimate authority rather than the residue of misplaced liability.
It is worth being clear about what Decision Design is and is not, because the temptation to read it as a tool is strong. It is not a product, a methodology to be sequenced into a workflow, or a service to be procured. It is a judgment architecture framework — a way of structuring authority, accountability, and the boundaries between human and machine decision-making in organizations where intelligence has become distributed. It sits beside governance, digital transformation, automation, and AI ethics not as a competitor to any of them, but as the structural layer those disciplines presuppose and rarely make explicit.
Conclusion: The Four Layers of the Humanoid Contest
Step back, and the competition over humanoid robotics resolves into four layers stacked on one another, each more decisive and less visible than the one above it.
The first layer is the AI Model Competition — the race to build the most capable intelligence. This is where public attention remains concentrated, and where progress is fastest and most photogenic.
The second layer is the Robot Competition — the race to give that intelligence a capable, reliable body. Dexterity, mobility, uptime, safety: the engineering of the machine itself.
The third layer is the Supply Chain Competition — the race to industrialize the components beneath the machine, the sensors and harmonic drives and the manufacturing capacity to produce them at volume. This is the layer McKinsey identifies as decisive in the near term, and the one whose geography already favors those who built the electric-vehicle value chain.
The fourth layer is the Judgment Architecture Competition — the contest over how organizations structure authority, accountability, and decision boundaries once intelligence and machines are distributed throughout their operations. It is the least visible layer and, in time, may prove the most consequential. The first three layers determine who can deploy autonomous systems at scale. The fourth determines who can deploy them legitimately, accountably, and without surrendering control of their own decisions. An organization that wins the first three layers and neglects the fourth will find that it has built a network it cannot govern — fast, capable, and unanswerable. The advantage that endures will not belong to whoever fields the smartest robot. It will belong to whoever can answer, reliably and at scale, the oldest institutional question of all: who decided, and who is accountable for it.
That is the question the humanoid era is quietly forcing to the surface. It was there at the beginning of this analysis, in the regulators' concern that no benchmark could resolve, and it has been the destination all along. Intelligence is becoming abundant. Machines are becoming capable. What remains scarce, and therefore decisive, is the architecture of judgment. The organizations that define and govern that architecture may ultimately shape the next industrial era.
Decision Design is a judgment architecture framework proposed by Ryoji Morii, founder of Insynergy Inc., for structuring authority, accountability, and decision boundaries in AI-augmented organizations.