Machines like us

A human hand rests on a robot’s forearm. Not on the emergency stop, where managers like to pose for photographs, but on the joint itself, steadying, guiding, correcting. The arm yields. It is compliant in the way a pupil is compliant – not limp, not inert, but receptive. The worker traces a small arc through the air. The machine follows. Together they hover above a workbench where an object, an awkward part, a fragile tool, a bowl of something viscous, waits to be handled correctly.

Nothing about this scene advertises a revolution. There is no new factory rising on the horizon, no headline about a “breakthrough” that will change everything “forever.” Yet something in the relationship has altered. For much of the industrial age the central drama of automation has been instruction – how to tell a machine what to do, and who gets to do the telling. Most robots, even sophisticated ones, have been trained in the narrow sense of the word – specified, scripted, parameterised. They have been obedient in the way clocks are obedient. And for that reason, they have lived where clocks live – in controlled environments, behind safety cages, inside carefully described worlds.

In the last few years, the centre of gravity has begun to move. Increasingly, robots are being taught physical tasks not by code, or simulation, or the consumption of endless video, but by something older than industry – demonstration. A person does what they already know how to do, fold, lift, wipe, pour, guide, and the machine records the movement as data. If it can generalise the demonstration (and this is the critical if), it can repeat the task tomorrow under slightly different conditions. A cup is an inch to the left, the surface is slicker, the weight is different. The robot adjusts, not because a programmer anticipated every variation, but because the learning system extracted a pattern from the demonstration and built a behaviour that can flex.

Call it imitation learning if you like, or learning from demonstration, or the milder phrase robotics researchers have used for decades, programming by demonstration. The label matters less than the political economy it implies. When robots can be trained through physical guidance, the scarce resource in automation shifts. For a century, the bottleneck was the ability to translate a messy human task into a formal instruction. That translation required engineers, integrators, process designers – people who could reduce work to a sequence that could be named, parameterised, and audited. The new bottleneck is closer to the shop floor – skilled movement itself. The robot is a pupil awaiting a competent body.

This is where the story becomes less about robotics and more about property, labour, and the state. If human movement can be captured, refined, replicated, and distributed, then physical skill begins to resemble a new class of economic asset, a form of intellectual property that is neither idea nor text, but motion. That prospect raises questions that are technological in origin but legal and moral in consequence. Who owns a demonstration? Who is paid when a robot repeats it? Who certifies that it is safe? What happens when a task taught in one jurisdiction is deployed in another, with different labour protections, different safety standards, different cultural norms of care?

The easy move is to treat this as one more chapter in the narrative of “AI disruption” – a future in which jobs are automated, productivity rises, inequality follows. But physical imitation has a particular sting. It does not simply replace labour, it extracts it, in the literal sense of taking what resides in the body, timing, pressure, posture, dexterity, and converting it into a replicable commodity. The old fear of automation was that the machine would take your job. The new fear is that the machine will take your job and keep a copy of your hands.

A shift without a name

It begins, as most large shifts do, quietly. The world of technology is loud, software launches, valuations, slogans, and this noisiness has made it hard to notice changes that are not easily rendered in screenshots. A robot being guided by a human hand is not as photogenic as a self-driving car or a chatbot composing sonnets. Yet it is a more fundamental claim on reality – a claim on friction, force, and the stubborn unpredictability of things.

The shift is methodological. Industrial robots have long been excellent at what economists politely call “structured tasks” – welding a seam on a predictable chassis, placing components on a board, painting a surface that does not move. Their success has come from precision and repetition. But outside the factory, in kitchens, hospitals, farms, construction sites, homes, work is not repetition so much as adaptation. The object is slightly different, the environment is messy, the conditions change, the correct action depends on touch as much as sight. There is no stable stage on which a robot can perform its choreographed routine.

This is why the “robot revolution” has always been partial. It transformed certain kinds of manufacturing and left much of the economy relatively untouched. Care work, domestic work, food preparation, repair, maintenance, much of logistics and construction – these have been resistant not because they are “too human,” but because they are too variable. A robot that must be explicitly programmed for every variation is economically useless in settings where variation is the norm.

Physical imitation learning offers a different route. Instead of requiring the world to be simplified so that robots can operate within it, it promises to make robots more fluent in the world’s complexity. The teacher does not explain in symbols, they demonstrate in movement. The robot records trajectories, forces, timing, and constraints. And in the ideal case, it builds a policy, a rule of action, that can be executed again under slightly different circumstances.

For strategists and policymakers, the implications are unnerving precisely because they do not come dressed as novelty. Teaching by demonstration resembles apprenticeship – a transmission of know-how through guided action. It is what humans have always done when they tried to teach something that could not be fully articulated. The difference is that the apprentice is an increasingly general machine whose memory is perfect, whose copying is cheap, and whose replication, once the skill is encoded, is essentially unlimited.

If earlier rounds of automation were tied to engineering capacity, this round is tied to the capacity to capture, curate, and govern embodied skill as data. The question becomes not only who builds robots, but who has access to demonstrators, who can create training pipelines, who can store and protect motion datasets, who can set standards that determine what counts as a “safe” or “certified” behaviour. And because these questions are political, not merely technical, they arrive in places the robotics industry would rather avoid – labour law, intellectual property, competition policy, privacy, and, inevitably, geopolitics.

From apprenticeship to Taylorism, and back again

“Tacit knowledge” is one of those phrases that academic writing has made both useful and dull. It names something obvious – much of what people know how to do cannot be cleanly described. A nurse lifting a patient understands, through experience, how to adjust to a body’s weight, a baker kneads dough with pressure calibrated to texture and temperature, a mechanic listens for a sound that is not in any manual. The knowledge is real, but it does not arrive as propositions. It arrives as trained attention and trained movement.

For most of history, economies were built on this kind of knowledge. Skills were transmitted through apprenticeship – the novice watched, copied, repeated, and was corrected. The body was the medium. The method was slow and intimate, but it worked. It produced craftsmen, carers, cooks, sailors, builders, people whose competence lived in their hands and nerves more than in their words.

Industrial capitalism, which required scale, found this model insufficient. It began the long project of extracting tacit skill from the body and embedding it in something that could travel – plans, blueprints, process charts, standard operating procedures. Frederick Winslow Taylor’s stopwatch, the Gilbreths’ motion studies, the assembly line – these were not only systems of production. They were systems of knowledge capture. They turned work into an object that could be measured, optimised, and reassigned. They wrested control of the labour process from workers and relocated it, first to managers and engineers, later to software.

Robotics inherited this logic. A robot’s competence has typically been designed elsewhere and delivered to the workplace as a finished script. Even when “learning” entered robotics, it often entered as a continuation of abstraction – learning from simulation, or learning from massive datasets of images and videos. These methods can work, and in certain domains they work brilliantly. But they have a particular blindness. They see the world without feeling it.

Most of what makes physical work difficult is haptic. Friction, compliance, resistance, micro-adjustment, the difference between a safe grip and a dangerous one – these are not easily inferred from pixels. When you teach a child to tie a shoelace, you do not show them a video. You guide their fingers. Physical imitation learning recognises this. It treats touch as information, and demonstration as a way of transmitting it.

In that sense, the technology is a return to history, though a return under new conditions. Apprenticeship comes back, but the apprentice is infinitely replicable. Taylorism comes back, but the motion study is now a dataset fed into a machine that can execute the motion at scale.

The political question is whether this new form of capture will repeat the old pattern – the extraction of skill from workers without corresponding compensation or control. The economic question is whether the return of apprenticeship will democratise automation, as enthusiasts promise, or simply create a more efficient mechanism for concentrating value.

What it means for a robot to learn by touch

It is tempting to treat “learning from demonstration” as an interface improvement – no more programming, just show the machine. In some cases, it will indeed look like that – a practical reduction in setup time, a faster way to configure a robot for a new task. But the deeper change is epistemic. Demonstration changes what a robot knows.

A programmed industrial robot knows a path in space. It moves from coordinate to coordinate. It is impressive when the world is constant and brittle when it is not. A robot trained by demonstration is, at least in principle, trained on a relationship – between movement and outcome, between force and compliance, between timing and success. It records not only where the hand went, but how it met resistance, not only the goal, but the method.

To make this possible, a modern robot needs a form of bodily awareness – sensing of force and torque, measurement of joint angles and velocities, and control mechanisms that allow safe compliance when a human guides it. It must be able to “yield” without collapsing, to register pressure without panicking, to replay a motion without injuring its surroundings. In practical terms this involves force/torque sensors, impedance and compliance control, and learning algorithms that can turn a handful of demonstrations into a generalisable behaviour.

The interesting point is not that these techniques exist, they have existed in some form for years, but that they are becoming economically and technically viable at scale. As hardware improves and learning systems become more data-efficient, physical demonstration begins to shift from laboratory curiosity to industrial practice. The training bottleneck moves. Tasks that used to require engineers to spend weeks scripting and debugging can be taught, in certain settings, by practitioners in hours.

This does not mean that programming disappears. It means that programming recedes from the interface. Engineers will still build the training systems, the safety envelopes, the simulation tools, the data pipelines, the standards. But the creation of task-level competence, the specific “how” of the work, begins to come from demonstrators – people who can do the task well, even if they cannot describe it in a formal language.

The effect is to turn skill into something like content. The analogy is imperfect but instructive. The web created an economy in which ordinary people, once merely consumers, became producers of digital artefacts, videos, posts, code snippets, tutorials, that could be monetised, copied, remixed. Physical imitation could create a similar dynamic for embodied labour. The “content” is movement – trajectories, force profiles, timing strategies, hand–object interactions. The “platform” is the repository where these motions are stored, curated, rated, licensed. The “creator” is the worker whose competence becomes a training dataset.

This is where breathless talk of “democratisation” becomes plausible, and where it becomes dangerous. If the new wealth of robotics lies not only in hardware but in libraries of teachable skills, then control over those libraries becomes a form of power. A country, firm, or platform that accumulates a dominant archive of high-quality demonstrations may control downstream automation markets as decisively as a software monopoly controls an app ecosystem. The industrial economy had its cartels of steel and oil. The digital economy had its monopolies of search and social media. The motion economy may produce monopolies of gesture.

Skill as data, and the birth of “skillware”

In the twentieth century, globalisation was measured in containers. A shipping box did not merely move goods, it standardised them. It created a system in which production could be fragmented across borders and coordinated by logistics. The result was not simply more trade but a different geometry of power – ports and shipping lanes became strategic assets, supply chains became sites of vulnerability, manufacturing moved to where labour was cheap and regulation lighter.

Something similar may be happening with skill. When a physical task can be recorded as data and redeployed anywhere, the movement itself becomes a tradable object. Demonstrations can be stored, refined, generalised, licensed, downloaded. The robot is no longer the only product. The training is.

There is a useful term here, if a slightly ugly one – skillware. Software is code that tells a machine what to do. Skillware is a package of movement competence, a taught behaviour that can be installed on a robot. It might be narrow (a particular fold, a particular grip) or complex (a sequence of actions in a care routine). Like software, it may have updates. Like software, it may have licensing terms. And like software, it may be distributed through platforms that take a cut, impose standards, and shape what is visible and valuable.

Imagine a marketplace, an app store for physical tasks. A bakery downloads a “croissant fold” routine taught by a master pâtissier. A warehouse downloads a “fragile packing” routine refined by expert workers. A care agency downloads a “patient transfer” routine certified for safety. This is an extrapolation from the logic of the technology. Once the marginal cost of replication approaches zero, the central economic question becomes – who owns the original?

This is where existing legal categories begin to fail. Intellectual property law is built around expressions (copyright), inventions (patents), and marks (trademarks). It is not built around kinaesthetic recordings. Labour law, in most jurisdictions, treats the output of work as the immediate product – the cleaned room, the assembled part, the delivered meal. It does not easily recognise the creation of a demonstration dataset as a separate, durable asset, despite the fact that, in a skillware economy, the demonstration may be worth more than the task itself. Competition policy is not yet tuned to the possibility of “gesture data monopolies” – firms that accumulate vast libraries of demonstrations and lock competitors out of automation markets downstream.

The stakes here are not confined to robotics. They are a preview of a broader problem – what happens when the body becomes an interface for data extraction, and when data extracted from the body becomes a form of capital?

One can already see the outlines. In the last two decades, the digital economy learned to convert attention into revenue. Social media platforms did not invent the human desire to look and be looked at, they built systems to capture and monetise it. The motion economy will attempt something analogous with skill. It will build systems to capture and monetise what people can do with their hands.

There is nothing inherently unjust in this. A society might decide that demonstrators should be paid as creators, that skillware should be licensed with royalties, that unions should negotiate over demonstration datasets, that public institutions should curate open libraries of essential tasks. Done well, this could elevate forms of labour that have long been underpaid and undervalued. But the default path, if policy remains passive, will be the path of extraction. Workers will be asked (or compelled) to demonstrate tasks as part of their employment. The demonstrations will be stored as corporate property. Robots trained on those demonstrations will displace workers. The value will travel upward.

The history of industrial capitalism suggests which path is more likely without intervention.

The labour question – when your hands become IP

The political romance of physical imitation learning is that it relocates agency. Instead of engineers dictating automation from above, practitioners teach machines from within. A baker becomes a trainer, a mechanic becomes an instructor, a cleaner becomes a knowledge producer. This is not merely a rhetorical elevation. If the robot learns through demonstration, the demonstrator is indeed the source of the robot’s competence.

But romance is not policy. The question is what happens to this agency once it enters the machinery of employment, contracts, and platforms.

Consider the act of demonstration. In a factory it might be framed as “training the system,” an additional task assigned to workers. In a care home it might be framed as “helping the robot learn the routine,” a minor inconvenience for staff. In a household it might be framed as “personalising your assistant,” a feature for consumers. In each case a person provides embodied knowledge, and in each case the resulting data may be valuable in ways the demonstrator does not fully control or even perceive.

If the value of skillware grows, the incentives will be clear. Firms will want to capture demonstrations at scale. Platforms will want to standardise and distribute them. Investors will want defensible assets – proprietary libraries of motion data, protected by contracts and technical barriers. Workers will want a share, or at least protections against exploitation. The friction here will not be merely economic but moral – the sense that something intimate, one’s way of doing a task, one’s trained competence, is being appropriated.

Labour movements have faced a version of this before. Taylorism was experienced by workers not only as speed-up but as theft – the theft of control over the labour process. When management turned work into a sequence of measurable motions, it did not only raise productivity, it reduced autonomy. Physical imitation learning risks repeating this dynamic in a new form. The worker does not simply lose control of the labour process, they supply the data that makes the loss possible.

This is why the seemingly technical question, who owns a demonstration?, is a labour question in disguise. One could imagine several regimes:

Employer ownership by default, where demonstrations created at work belong to the firm, like any other output. This is the path of least resistance, and it will produce predictable outcomes – rapid corporate accumulation of skill libraries and rapid worker dispossession.

Creator-like rights for demonstrators, where the act of demonstration is recognised as producing a durable asset and is compensated accordingly, through royalties, licensing, or equity-like claims.

Collective governance, where unions or worker councils negotiate over demonstration datasets, including rules for consent, compensation, reuse, and deletion.

Public interest regimes, where certain classes of demonstrations, especially in essential services like care, are treated as public goods, curated in open libraries with safeguards, to prevent monopolisation.

Each regime implies a different future. The most plausible future, absent policy, is the first. Firms will capture skillware as an intangible asset, and workers will be told they were paid already. The second, third, and fourth futures require law – new categories of rights, new contract norms, new institutions.

There is also the question of what happens to workers who become trainers. A new occupational category emerges – the machine demonstrator, the task instructor, the quality assessor of learned behaviours. These roles can be dignified and well paid, or precarious and invisible, depending on how they are structured. The gig economy offers a cautionary tale. Platforms often frame work as flexible and empowering, then they use that framing to deny workers the protections that make work bearable. A skillware economy could do something similar, turning skilled demonstrators into “contributors” paid per dataset, competing in a global marketplace where the terms are set by platforms.

If that happens, physical imitation learning will not democratise automation. It will proletarianise training.

The state and the geopolitics of teachability

The most seductive claim about physical imitation learning is that it lowers barriers. If robots can be taught by demonstration, then automation becomes possible in settings that previously lacked the engineering capacity to program it. Small firms could train robots without hiring a team of integrators. Emerging economies could leverage abundant skilled labour without needing a large software sector. Vocational expertise could become a comparative advantage.

This is not implausible. Yet it must be treated with caution. Democratisation of training is not the same as democratisation of ownership. A country may have millions of skilled workers and still find itself buying robot “skills” from foreign platforms. A small business may be able to teach a robot and still find that the terms of licensing and updates are controlled elsewhere. Lowering the barrier to entry for learning does not automatically lower the barrier to power.

Even so, the geopolitical implications are real. For decades, national competition in automation has centred on hardware manufacturing, semiconductor capacity, and engineering talent. Physical imitation learning shifts the emphasis toward workforce teachability, the ability to digitise embodied skill systematically. Countries with robust vocational systems, strong craft traditions, or large informal economies may find themselves unexpectedly advantaged if they build the infrastructure to capture and govern demonstrations.

This suggests a new kind of industrial policy. It will not be enough to subsidise robotics firms or fund AI research. States may need to invest in:

Skill capture infrastructure – training centres, sensor-equipped workplaces, standardised tooling for recording demonstrations safely and consistently.

Governance frameworks – rules for consent and compensation, standards for safety and auditability, mechanisms for portability and interoperability.

Public repositories – open libraries of essential tasks, especially in domains like care and public services, to prevent monopolisation of socially critical skillware.

Education for “machine pedagogy” – training workers not to code, but to teach, how to decompose tasks, provide high-quality demonstrations, recognise unsafe generalisations, and supervise robots effectively.

These investments would treat embodied knowledge as a national asset, not in a chauvinistic sense but in a practical one. Nations already compete over data and standards. Motion data will be no different. The state that understands this early may become a hub not just of robot manufacturing but of robot competence.

There is, however, a darker version of the same logic. States might also treat demonstrations as resources to be extracted and controlled, building national archives of skill in ways that resemble surveillance. The line between skill capture and worker monitoring is thin. If motion sensors record demonstrations, they can also record performance. A system built to teach robots can be used to discipline humans. The labour politics here will be delicate.

Safety, culture, and the problem of black-box movement

The most comfortable fantasies about robots are those in which machines behave like appliances – reliable, predictable, safe. Physical imitation learning unsettles this. A robot trained through demonstration may perform a task beautifully in the conditions it has seen, and dangerously in those it has not. Unlike code, demonstrations are hard to audit. Unlike written procedures, they are not easily reviewed by inspectors. A learning system may generalise in ways that are difficult to anticipate.

This creates a regulatory problem. Most automation regulation assumes machines are deterministic – their behaviours can be traced to specifications and verified against standards. A robot trained through imitation is not entirely opaque, but its competence emerges from data and learning dynamics, not from a single auditable script. Regulators will need new tools – provenance requirements for training data, audit trails for retraining, certification regimes for learned behaviours, and standards for “when not to act.”

Safety is not the only issue. Culture is another.

Physical demonstrations carry norms. A caregiver’s movements encode assumptions about dignity and privacy. A service worker’s gestures encode social expectations. A “correct” way to handle a body in one culture may be inappropriate in another. A robot trained in one jurisdiction may be deployed elsewhere. When that happens, the robot brings its learned norms with it. This is not a speculative worry. We have already seen how systems trained on narrow datasets fail when deployed broadly. In physical tasks, the consequences can be more than inconvenience.

The domestic sphere intensifies these concerns. A robot taught in a home will, by definition, learn intimate routines – how a family handles food, how an elderly person is lifted, how children are cared for. That data is as valuable as it is sensitive. It may reveal health conditions, habits, vulnerabilities. Privacy law has barely begun to grapple with domestic sensor data from smart speakers and cameras. Motion data raises the stakes again. A demonstration recorded in a home should not become a platform asset by default. Households should have meaningful rights – to delete, to limit reuse, to know where the data travels, to refuse secondary use.

Finally, there is the problem of hyper-replication. Once a task is taught, it can be repeated at scale. That is the point. But it is also the source of displacement. One excellent demonstration can produce millions of competent executions. Efficiency looks like progress in economic statistics. To the worker whose livelihood depended on the task, it looks like erasure.

The challenge is not to prevent replication but to socialise its benefits. Automation has repeatedly increased productivity without increasing worker welfare proportionally. There is no reason to assume this wave will be different on its own. If physical imitation learning becomes a new productivity layer, then policy must decide how the gains flow – into wages, into public services, into reduced working hours, into worker ownership, rather than only into platform rents.

A politics for embodied intelligence

We are used to thinking of the knowledge economy as disembodied – code, information, data in the cloud. Physical imitation learning forces a different recognition. A great deal of economically valuable knowledge has always been embodied. The novelty is that it can now be captured and scaled.

This is why the appropriate response is not simply “more innovation,” nor simply “more regulation,” but something closer to constitutional design – a governance regime for embodied intelligence.

Such a regime would begin with ownership and compensation. If demonstrations are the fuel of the new automation, then demonstrators are not peripheral labour. They are creators of capital. The law should reflect that – through rights of attribution, compensation models (including royalties or collective licensing), and enforceable consent standards.

It would also need standards for safety and auditability that recognise how learning systems behave – not just pass/fail tests of tasks in ideal conditions, but stress-testing of generalisation, requirements for human override, clear escalation pathways, and certification processes especially strict in safety-critical domains like healthcare and construction.

Competition policy would need to anticipate concentration. The future may belong to platforms that curate and distribute skillware. The question is whether they will become the equivalent of app stores, useful but extractive, or the equivalent of public utilities, interoperable, regulated, accountable. Data portability and interoperability may matter as much for motion datasets as they have for social media and finance.

Education policy, too, must shift. The last decade’s mantra, teach everyone to code, was not wrong so much as incomplete. The coming decade may require teaching people to teach machines – a curriculum of task decomposition, demonstration quality, robot supervision, and safety awareness. If done well, this could elevate vocational expertise rather than dismiss it as pre-digital residue. The skilled worker would no longer be only a cost. They would be an asset in the literal sense – a producer of transferrable competence.

And there should be public investment in open skill libraries for essential tasks. The idea is the prevention of monopoly in domains where monopoly would be morally intolerable. A society that allows a handful of firms to own the “standard” datasets for care routines, basic cleaning, household assistance, would be handing private platforms a form of infrastructural power.

One could summarise the entire issue in a single question – what happens when the most valuable dataset is a person who knows how to do something well?

In the old economy, value moved in parts. In the digital economy, it moved in information. In the coming economy, it may move in motion. If that is true, then the politics of the twenty-first century will include a new dispute – not only over who owns factories and code, but over who owns the captured competence of human bodies.

The final choice is not whether robots will learn from us. They already are. The choice is whether we build a world in which that learning becomes another route for extraction, or a route for shared prosperity, dignity, and control over the skills that have always been the hidden infrastructure of civilisation.