When Intelligence Acquires a Body: Will Robotics Follow the AI Explosion?

Editorial image generated by the author.

AI may be giving robotics a scalable intelligence layer. But intelligence can spread like software; robots still have to be built.

For most of the recent AI boom, the revolution has remained strangely immaterial. Robotics is the obvious missing half. Models can write software, analyse documents, generate images, design molecules, translate languages and increasingly operate computers, yet most of the physical world still functions much as it did before ChatGPT appeared. Warehouses contain human workers. Builders carry materials around construction sites. People clean hotel rooms, stock supermarket shelves and load washing machines.

The usual explanation is straightforward. Manipulating information is easier than manipulating the physical world. A better AI model can be copied to another server almost instantly; a better robot must be manufactured. Software can attempt a task a thousand times without breaking anything. A robot dropping a glass, damaging a machine or colliding with a person is a different proposition.

That has encouraged us to treat AI and robotics as separate technological stories: intelligence might continue its rapid ascent while robots follow the familiar, slower path from one carefully engineered industrial application to another. There is now a plausible reason to doubt that assumption. The more interesting possibility is a robotics explosion in two stages: robotic intelligence may begin improving at something approaching software speed, while physical deployment follows on the slower timetable of factories, supply chains and capital investment.

If that is what is happening, the important breakthrough is not that humanoid robots can run, dance or perform increasingly elaborate demonstrations. It is that the intelligence required to control machines may be becoming reusable across different bodies and tasks. That would not turn robotics into software. Atoms remain stubbornly different from bits. It could, however, remove one of the constraints that has kept general-purpose robotics perpetually just over the horizon.

When Robotic Intelligence Becomes Reusable

Traditional industrial robots can be extraordinarily capable. A robotic arm on an automobile production line may perform the same operation thousands of times with a speed and precision no human worker could match. Its intelligence is usually narrow: change the object, workstation or process and engineers may have to recalibrate sensors, redesign motion paths, replace fixtures and rewrite software. The model has nevertheless produced an enormous industry. According to the International Federation of Robotics’ World Robotics 2026 report, 603,000 industrial robots were installed in 2025, bringing the global operational stock to about 5.08 million. China installed 354,000 of them, accounting for 59% of global deployments.

What this model has not produced is a genuinely general-purpose robot. If every new task requires substantial engineering work, there is no equivalent of installing another application on a computer. A machine designed to pack one particular product into one particular box may be extremely valuable, but one that can encounter unfamiliar products, receive instructions and work out how to pack them itself belongs to a different economic category.

Recent work in robot learning suggests that the cost of moving from one task or body to another may be falling. In a recent essay, robotics researcher Eric Jang argues that the field may be entering what he calls a “smooth capability exponential”. His most interesting observation concerns training data. Instead of requiring enormous amounts of experience collected from the exact machine that will eventually perform a task, a growing share of physical competence may be learned from broader sources: human video, egocentric recordings, gripper demonstrations, simulation and data gathered from other robot designs. Much smaller quantities of high-quality data from the final machine can then adapt that knowledge to a particular body.

Jang describes a data pyramid, with enormous quantities of relatively hardware-independent experience at the bottom and relatively little machine-specific training at the top. Google DeepMind reports an early example of this pattern. Its Gemini Robotics On-Device 2 can be adapted to substantially different bi-arm robot configurations using only a few hours of data and typically fewer than 200 examples. That is a manufacturer-reported result from a model still offered through limited access, not proof that arbitrary robots can acquire arbitrary skills from 200 demonstrations. It nevertheless suggests that learned physical competence is becoming less tightly coupled to any one machine.

NVIDIA is pursuing the same broad direction. Its GR00T 1.7 model was pretrained on roughly 32,000 hours of real robot demonstrations and human egocentric data together with about 8,000 hours of simulated rollouts and demonstrations, before being adapted to particular tasks and embodiments.

The emerging model is therefore less new robot + new task = new engineering project and increasingly general physical model + embodiment-specific adaptation = many possible tasks. Robotics does not need complete independence from hardware for this to matter. If the marginal cost of adapting an existing body of learned physical knowledge to another machine keeps falling, the field begins to inherit one of the foundation model’s most valuable properties: knowledge acquired in one context becomes reusable elsewhere.

That creates feedback loops traditional robotics never really possessed. Better models make robots useful for more jobs; more useful robots justify greater deployment; more deployment produces more real-world data; and that data feeds the next generation of models. AI also accelerates software development and simulation, while larger markets give component suppliers stronger incentives to improve motors, gearboxes, sensors and actuators.

China makes the ecosystem effect particularly visible. It combines a vast electronics and manufacturing base with heavy investment in robotics. Reuters reported in September 2026 that Chinese companies accounted for roughly 95% of global humanoid shipments in 2025, even though many of those machines remain experimental rather than economically productive. The exact share matters less than the industrial system forming behind it. Motors, batteries, sensors, precision components and control electronics can support many different kinds of machine. Even if some current humanoid designs prove to be evolutionary dead ends, the factories and suppliers created around them need not be.

The Reliability Wall

A robot that can perform a task is still a long way from one that can be trusted to perform it throughout a working day. Google’s own published results illustrate that gap unusually well. Its Gemini Robotics 2 system can control a humanoid across its whole body while combining movement with manipulation, an impressive research result. In Google’s evaluations on the Apollo humanoid, however, picking an object from a table succeeded 68.4% of the time, picking from a shelf 76.3%, and picking something up from the floor 45.7%. Multi-finger manipulation was harder still: tying a rubbish bag reached 44%, while a dustpan task reached 32%.

These numbers are not evidence that the technology is failing. A fairly general model is attempting physical tasks that would until recently have required much more specialised engineering. They do show why a spectacular demonstration and an economically useful worker are different things.

Small failure probabilities become particularly unforgiving over longer sequences. As a deliberately simplified illustration, imagine a household chore containing thirty independent actions. At 99% reliability per action, the probability of completing the entire sequence without a single failure is only about 74%. At 98%, it falls to roughly 55%. Around 99.9% per step raises the overall figure above 97%.

Real robotic systems are more complicated. Failures are not independent, some errors barely matter, and capable systems can detect problems and retry. Google itself reports progress in self-correction and longer multi-step sequences. The arithmetic nevertheless captures the underlying problem: a machine can look astonishingly competent for five minutes and remain maddeningly unreliable over an eight-hour shift.

The same tension appears in China’s rapidly expanding humanoid industry. A Reuters investigation found sophisticated hardware alongside machines that remained slow, awkward and heavily dependent on training for practical factory work. Conventional industrial equipment still outperforms humanoids for many tasks that can already be automated. Over the next few years, the most revealing robotics metric may therefore be neither walking speed nor the sophistication of a selected demonstration. It may be the failure rate during hours of unsupervised work.

Atoms Do Not Scale Like Software

Even if the intelligence problem improves faster than expected, another constraint remains. A better model can be copied almost without limit; a better robot cannot. Suppose somebody released an extraordinary general-purpose robotic model tomorrow. The software might spread rapidly, but the machines running it would still require factories, motors, precision bearings, gearboxes, batteries, sensors, processors, wiring, testing, logistics and maintenance.

Hardware economics are improving, but unevenly. McKinsey estimates that actuation accounts for roughly 40% to 60% of the bill of materials of a humanoid robot. Batteries and computing hardware can draw on enormous industries built for electric vehicles and electronics. High-performance robotic actuators, tactile systems and some other specialised components remain much smaller markets.

This produces two different curves. Robotic capability may increasingly behave like software; robotic deployment remains an industrial process. The first could improve surprisingly quickly. The second is more likely to resemble the adoption of automobiles, industrial machinery, solar panels or electric vehicles: very rapid once the economics become compelling, but still constrained by factories, capital expenditure and supply chains.

The difference changes the timing rather than eliminating the transformation. Millions of people can gain access to a better AI model within days of its release. If robotic intelligence improves equally quickly, most people may initially experience the change only indirectly. The software could be ready long before enough physical machines exist to embody it.

Where Physical AI Becomes Economical First

Jang predicts that the first two or three general-purpose home robots could reach consumers around October 2027. He does not imagine them immediately replacing domestic labour. The first machines might perform simple pick-and-place jobs, fold some laundry and offer enough novelty to attract a few thousand enthusiasts before later generations become genuinely useful.

That is plausible as a path to an early consumer product. It is less obviously the path to economic transformation. A household considering an expensive robot asks what it can do that justifies the purchase. A manufacturer asks whether the machine can operate for two shifts a day while reducing labour, downtime or retooling costs.

Factories, warehouses and other semi-structured environments are also friendlier places for early deployment. Floors are predictable, objects can be standardised, lighting can be controlled and dangerous areas can be isolated. Tasks can be chosen precisely because machines already perform them reasonably well. A home contains stairs, loose clothes, glassware, liquids, pets, children, cables and thousands of objects placed differently from one day to the next. Domestic robotics may eventually become an enormous market; it may also be one of the hardest environments to solve.

Nor does the machine have to look human. Humanoids have an obvious advantage because civilisation has already constructed an enormous environment around human dimensions. Stairs assume legs. Shelves assume a certain reach. Doors assume hands. Workstations, tools and vehicles were designed for bodies roughly like ours.

The human form is nevertheless not optimal for every job. A warehouse machine may work better on wheels. An agricultural robot might benefit from several arms. A construction machine may use tracks or interchangeable tools. A stationary arm will continue to outperform an expensive walking robot wherever the work never requires walking.

The deeper technological shift is therefore not humanoids as such. It is general-purpose physical AI: reusable models that can understand instructions, recognise unfamiliar objects, reason about physical consequences and transfer skills between different bodies. If that becomes practical, embodiment increasingly becomes an engineering choice rather than the place where all the intelligence has to be rebuilt from scratch.

When Robots Start Expanding Productive Capacity

The economically important threshold is also lower than “human-level robotics”. A warehouse robot does not need to understand literature, an agricultural machine does not need to drive a taxi, and a construction robot does not need to cook dinner. Nor does a useful machine necessarily have to outperform the best human worker. It has to perform enough valuable tasks, for enough hours, at sufficient reliability, for its lifetime cost to beat the available alternative.

Industrial automation already works this way. Factories did not wait for general artificial intelligence before adopting robots; they automated whichever parts of production offered a clear return on capital. General physical AI changes the calculation by widening the set of tasks that the same investment can perform.

A robot that performs one operation is machinery. A robot whose software allows it to perform dozens of different operations begins to resemble a platform. A company may be reluctant to invest heavily in a machine designed for a process that could disappear in three years. A machine that can be reassigned as requirements change has a different economic life.

Beyond that lies a more speculative possibility. AI has so far automated parts of cognition while leaving the expansion of physical productive capacity largely untouched. A model can design a factory but cannot pour its foundations. It can optimise a solar installation but cannot mount the panels. It can design a better motor but cannot operate the machinery that manufactures it.

Robotics could begin closing that loop. The first machines will perform individual jobs: moving containers, sorting goods, assembling components and inspecting infrastructure. As capability broadens, robots could participate in more stages of production. Factories producing robots can themselves use robots. Automated mining and logistics can provide materials. Robotic construction could help expand factories, warehouses and energy infrastructure.

If automation eventually penetrates capital-goods production deeply enough, it stops affecting only the production of things we consume and begins affecting the production of productive capacity itself. A second feedback loop then becomes possible: better AI → better robots → greater productive capacity → more robots, energy and computing infrastructure → better AI.

Calling that outcome inevitable would be reckless. Every arrow hides difficult engineering, capital and energy constraints. Robots that merely deliver parcels or clean floors do not create recursive industrial growth. The loop becomes powerful only if automation spreads into manufacturing, construction, energy and the production of machinery itself. That is also why robotics could eventually matter more economically than the current generation of AI applications. The AI boom makes some forms of cognition dramatically cheaper; physical AI could begin acting on the constraints that remain after cognition is no longer scarce.

Intelligence First, Factories Second

The best evidence that a robotics explosion has begun will probably be mundane rather than spectacular. Watch how much robot-specific data is required to teach a machine something new. Watch whether models continue to transfer between substantially different bodies. Watch failure rates over hours of autonomous operation rather than success in selected demonstrations, and whether companies can move the same machines between jobs without calling in a robotics engineering team.

Above all, watch the economics. The decisive transition comes when companies routinely buy relatively general robots because their cost per productive hour is competitive across a growing collection of tasks. At that point robotics has moved beyond impressive research and into ordinary capital investment. The installed base would still take years to expand—factories cannot be summoned with an API call—but the direction would become much harder to reverse.

It is therefore too early to say that the AI explosion will naturally be followed by a robotics explosion. Reliability, dexterity, energy consumption, safety, cost and manufacturing capacity remain substantial obstacles. The argument is nevertheless considerably stronger than it was only a few years ago. The decisive development is not another humanoid walking faster or another polished factory demonstration. It is the possibility that a growing body of knowledge about the physical world can be learned once and reused across machines.

If that continues, robotics will have acquired something it has historically lacked: a scalable intelligence layer. We may then pass through a peculiar period in which robotic intelligence becomes astonishingly capable while economically useful general robots remain relatively scarce. That gap could look like another failure of robotics to fulfil decades of promises. It may instead be exactly what a transition from software-speed intelligence to factory-speed deployment should look like.

The intelligence can arrive first. The factories have to follow.

If they do, the AI revolution crosses an important boundary. Machines cease to be confined primarily to manipulating information and begin acting directly on the physical systems through which economies produce food, buildings, energy, machinery and eventually more machines.

The interesting question will then no longer be whether AI can replace particular cognitive tasks. It will be what happens when scalable machine intelligence acquires hands.