Abstract
Artificial intelligence may become capable of automating cognitive work substantially faster than industry can manufacture and deploy robotic bodies. This asymmetry could create an embodiment lag, an interval during which artificial cognition is abundant while artificial locomotion, manipulation, and physical presence remain scarce. The present article develops a labor-market framework for this transitional period. Anticipatory robotification is defined as the redesign of jobs, workflows, facilities, and production systems around expected robotic labor before enough robots exist to occupy the resulting positions. These robot-oriented positions are termed robot slots. When people temporarily occupy them because robotic capital is unavailable, they engage in reverse robotic substitution: humans substitute for robots within processes whose intended long-term workers are machines.
The likely labor-market expression of this arrangement is an embodiment gig economy in which people supply physical presence, mobility, dexterity, and force to AI-directed production systems on a shift-by-shift or task-by-task basis. Artificial intelligence can provide navigation, sequencing, occupational knowledge, quality control, and supervision through mobile devices, augmented-reality glasses, cameras, and wearable interfaces. This externalization of procedural knowledge may sharply lower entry barriers, allowing displaced knowledge workers and other participants to enter unfamiliar physical workplaces without making conventional career commitments. Human demand within robot-native production may consequently follow a hump-shaped trajectory: limited during early robotification, rising as robot-oriented processes proliferate faster than robot production, and declining when robotic fleets become abundant and facilities cease maintaining human-compatible operating modes.
The framework predicts the emergence of human embodiment tasks, portable embodiment credentials, substrate-flexible production systems, machine-derived performance standards, and new conflicts over surveillance and robot-training data. It also identifies policy requirements involving employment status, occupational safety, algorithmic transparency, worker control over embodied data, and compensation when human demonstrations contribute to robotic replacement. Human manual labor may therefore persist for years inside an economy already designed for its eventual elimination.
Keywords: humanoid robots, future of work, technological unemployment, gig economy, algorithmic management, embodied AI, augmented reality, robotification, labor platforms, human-robot interaction
1. Introduction
The automation of cognitive and physical labor is unlikely to proceed at the same rate. An artificial intelligence system can be deployed to additional users by allocating compute, integrating software, and expanding data-center capacity. A physical robot requires a body composed of actuators, sensors, processors, batteries, structural components, wiring, joints, hands or end effectors, safety systems, and communications hardware. It must then be assembled, tested, transported, installed, maintained, repaired, powered, and integrated into a particular workplace.
This difference could divide the transition to advanced automation into two partially separated technological waves. The first wave would make artificial cognition increasingly abundant. The second would create artificial embodiment at industrial scale. If the first wave advances substantially faster than the second, the economy may become capable of automating much of the planning, communication, analysis, coordination, and procedural knowledge involved in work while remaining unable to automate a comparable volume of physical execution.
Current occupational exposure estimates are consistent with such an asymmetry. The International Labour Organization has found the highest generative-AI exposure among clerical and strongly digitized occupations, while work dominated by manual and physical tasks generally has lower direct exposure to present generative systems. The ILO emphasizes transformation rather than inevitable elimination, but its findings indicate that digital work is technically accessible to generative AI earlier than many forms of embodied work.
The physical capital base remains comparatively limited. The International Federation of Robotics estimated that 4,663,773 industrial robots were operating worldwide at the end of 2024, with 542,076 installations during that year. These figures encompass established industrial machines rather than a fleet of general-purpose humanoids. The IFR characterized sales of multipurpose humanoids as very limited, although practical trials and substantial investments were underway.
The disparity is large enough to motivate a transitional theory of labor. Even if artificial intelligence becomes capable of controlling a competent humanoid worker, the existence of the control system does not instantly produce millions of robot bodies. Software capability and embodied capacity are separate variables. Intelligence may become available before sufficient machinery exists to enact it.
The central proposal of this article is that companies will begin reorganizing work around robotic labor before robotic bodies become abundant. They will identify tasks best performed by robots, standardize those tasks, develop machine-readable workflows, alter facilities, and treat robotic execution as the long-term production norm. Yet many of the positions created by that reorganization may initially remain vacant from the machine side. Human beings will then occupy jobs intended for robots.
This creates an inversion of the familiar automation narrative. Ordinarily, a robot enters a workplace and substitutes for a human. During the proposed transition, a human enters a robot-oriented workflow and substitutes for an unavailable robot. The person may be directed by artificial intelligence, trained through robotic demonstrations, evaluated against machine-derived performance models, and retained only until robotic supply catches up.
The likely institutional form is an on-demand market for embodiment. People may sign into an application, select a nearby physical shift, arrive at a warehouse or production site, put on an instructional device, and perform a sequence of actions generated by an AI management system. The worker need not master an enduring occupation because the occupational knowledge remains in the platform. The person contributes a temporarily available biological body.
This article develops three principal concepts:
- Anticipatory robotification, in which economic processes are redesigned around robotic labor before robotic labor is available in sufficient quantity.
- Reverse robotic substitution, in which people temporarily occupy positions whose intended long-term occupants are robots.
- The embodiment gig economy, in which platforms coordinate short-duration access to human physical capabilities for AI-directed or robot-oriented production.
The framework is conditional rather than chronological. It does not require that most white-collar jobs disappear by a specified year, that humanoid robots become the dominant machine morphology, or that universal basic income follow any predetermined schedule. It describes the likely consequences if cognitive automation becomes broadly deployable before artificial embodiment reaches comparable scale.
2. The Embodiment Lag
2.1 Artificial cognition and artificial bodies
Artificial cognition and artificial embodiment have different replication functions. An additional AI worker requires a share of computational infrastructure, but it does not necessarily require a new factory-sized supply chain for each instance. Multiple agents may also operate sequentially or concurrently on shared hardware, and software improvements can propagate across a fleet almost immediately.
An additional robot requires a separately instantiated physical system. Even where production is highly automated, every unit consumes materials, components, assembly time, quality-control capacity, logistics, and maintenance resources. The hardware may be improved through software updates, but an update cannot repair an undersized motor, add a missing tactile sensor, replace an unreliable joint, or create another body.
The embodiment lag can therefore be defined as:
The interval during which artificial cognition is sufficiently capable and abundant to direct or automate large portions of economic activity, while artificial bodies remain too scarce, expensive, unreliable, or geographically unavailable to execute the corresponding volume of physical work.
The lag encompasses more than shortages of raw materials. It includes limitations in factory construction, component yields, actuator production, battery manufacturing, supplier qualification, maintenance labor, charging infrastructure, capital financing, safety certification, and organizational integration.
The early scaling of humanoid production illustrates these demands. Figure reported in April 2026 that its BotQ facility had demonstrated a one-robot-per-hour production cycle, qualified hundreds of suppliers, built more than 9,000 actuators across multiple designs, and established more than 50 in-process inspection points. These are meaningful advances, but they also reveal that robot replication is an industrial production problem rather than a software-copying problem.
The one-robot-per-hour milestone would produce only thousands of machines annually if sustained on a single line. Figure had previously described its first-generation manufacturing line as capable of producing up to 12,000 humanoids per year. Even rapid multiplication of such facilities would require time to create a robotic capital stock comparable to the hundreds of millions of people engaged in physical work.
Supply-chain concentration could further constrain expansion. Permanent magnets containing neodymium, praseodymium, dysprosium, and terbium support precise, efficient motion in many motors. The International Energy Agency reports that robotics and automation will contribute to increasing magnet demand and that China accounted for 91 percent of refined magnet rare-earth output in 2024, with an even larger share of permanent-magnet production. Batteries, semiconductors, bearings, precision gear systems, copper, and specialized manufacturing equipment introduce additional dependencies.
2.2 Capability is not capacity
It is useful to distinguish robotic capability from robotic capacity. Capability concerns whether a robot can perform a task under the relevant conditions. Capacity concerns how much total work the available fleet can perform.
A breakthrough in dexterous manipulation could make a robot technically suitable for ten million positions without creating ten million units. Similarly, a general-purpose control model could be downloaded to every compatible robot while the number of compatible robots remains low. An economy can consequently possess the technical solution to a labor problem without possessing enough physical capital to implement that solution universally.
Early deployments provide a glimpse of this distinction. Figure reported that its second-generation robots operated for more than 1,250 hours at a BMW plant, loaded more than 90,000 parts, and contributed to the production of more than 30,000 vehicles during an 11-month deployment. The task was a constrained sheet-metal loading operation with explicit requirements for cycle time, placement accuracy, and human intervention. It demonstrated industrial usefulness while also generating hardware reliability data that informed the next robot generation.
The same company later demonstrated a four-minute autonomous sequence containing 61 room-scale locomotion and manipulation actions. Figure described these results as early, which is an important qualification. A system may display rapidly improving dexterity while remaining far from the uptime, maintainability, cost, and manufacturing volume required for economy-wide substitution.
2.3 Humanoids as compatibility machines
Humanoid morphology has a particular advantage during the first stage of physical automation. Buildings, tools, stairs, shelves, vehicles, work surfaces, handles, containers, and safety procedures were designed around human dimensions. A machine with approximately human reach, locomotion, and manipulation can enter those environments without requiring a complete reconstruction of the facility.
Apptronik has explicitly presented this compatibility as an advantage of its Apollo platform. Its Mercedes-Benz pilot focused on delivering assembly kits and moving containers within manufacturing spaces originally designed for people. The company argued that humanoid robots can reduce the need for full-scale facility redesign.
Humanoid compatibility does not imply that the mature robot economy will remain humanoid. Once an entire process is designed for machines, human anatomy may become an unnecessary constraint. A robot-native warehouse could use ceiling manipulators, autonomous carts, mobile shelves, conveyors, fixed arms, drones, and specialized grippers. A robot-native construction process could employ modular components and large nonhuman manipulators. A robot-native kitchen might use enclosed ingredient systems rather than bipedal cooks.
The embodiment lag should therefore be interpreted broadly. It concerns shortages of machine bodies and physical automation systems, including humanoids, mobile manipulators, specialized robots, and automated infrastructure. Humanoids are especially relevant to the proposed transitional period because a humanoid-compatible workflow is more likely to remain temporarily occupiable by people.
3. Anticipatory Robotification
Robotification can refer broadly to the growing organization of production around robotic labor. The more specific process proposed here is anticipatory robotification:
Anticipatory robotification is the redesign of jobs, tools, workflows, facilities, and production systems around expected robotic execution before sufficient robotic capacity exists to perform all of the resulting work.
The defining feature is a temporal mismatch between design and deployment. Companies begin treating the robot as the normative future worker while continuing to employ people because the intended machines are unavailable.
3.1 The robot as the normative worker
Every production system embodies assumptions about its worker. Shelf height, handle shape, walking distance, instruction format, cycle time, temperature, lighting, noise, and permissible force all reflect the capabilities of the expected operator. In a human-oriented system, the person is the default worker and the robot must be adapted to human infrastructure.
Anticipatory robotification reverses that relationship. The robot becomes the default worker in the production model, even when a person still occupies the station. The workflow may be represented as a machine-readable task graph. Objects may be standardized for robotic perception and grasping. Workstations may include fiducial markers, machine-readable labels, fixed presentation angles, automatic quality checks, and spatial arrangements chosen for robotic reach.
A position within this system can be called a robot slot:
A robot slot is a role in a production process whose tasks, interfaces, performance assumptions, or long-term economics are designed primarily for robotic execution.
A robot slot need not be filled by a humanoid. It may eventually be occupied by a mobile manipulator, fixed arm, autonomous cart, or integrated machine system. Its defining feature is that the production process treats machine execution as the desired equilibrium.
3.2 Why firms would robotify before obtaining robots
Several incentives could encourage anticipatory design. First, standardizing a workflow for robots can improve efficiency even while people perform it. Clear task decomposition, optimized routing, machine-readable inventory, automated verification, and real-time scheduling may reduce errors and training costs.
Second, anticipatory design creates deployment readiness. A company that waits for robots to become cheap before reorganizing its operation may lose time adapting facilities, data systems, and procedures. A company that develops robot-compatible workflows in advance can install machines as soon as units become available.
Third, robot-oriented standardization can make human labor more interchangeable. When the production system contains the procedural knowledge, workers require less site-specific expertise. This lowers the cost of turnover and allows firms to draw from a larger temporary labor pool.
Fourth, human performance can generate the data needed to complete robotic automation. People working inside instrumented robot slots expose edge cases, unusual object states, environmental variation, and recovery strategies. Their behavior can be used to improve the robotic policy that will eventually occupy the position.
3.3 Substrate-flexible production
During the embodiment lag, firms may benefit from building systems that can accept either a person or a robot. This arrangement can be called substrate-flexible production:
Substrate-flexible production is a production architecture in which the same economic objective or task graph can be assigned to either a biological or artificial body, with the execution policy adapted to the capabilities of the selected embodiment.
The central management system might represent a task as “move component A from storage position B to fixture C and verify orientation.” A robot receives coordinates, trajectories, grip parameters, force limits, and machine-state signals. A person receives visual directions, spoken instructions, warnings, and confirmation prompts. Both executions update the same production record.
Substrate flexibility provides insurance. Humans can cover robot shortages, breakdowns, unusual demand, maintenance periods, remote locations, or tasks that remain outside the machine’s reliable frontier. Robots can cover repetitive, hazardous, undesirable, or overnight shifts. The allocation system can assign each task to the available embodiment with the lowest expected cost.
The human version of a robot-native task requires a distinct human mode. Human mode translates the task into biologically appropriate movements, limits, and timing. It may reduce loads, slow cycle times, insert breaks, alter routes, restrict access to high-speed zones, and present additional explanations.
Without a human mode, substrate flexibility could become unsafe. A machine policy may assume unlimited repetition, precise timing, no boredom, no social needs, and different tolerances for heat, noise, acceleration, or impact. Human compatibility must be engineered rather than presumed.
4. Reverse Robotic Substitution
The traditional formulation of automation describes machines as substitutes for workers. Reverse robotic substitution describes the transitional inversion:
Reverse robotic substitution occurs when a person temporarily performs a position designed or designated for a robot because the required robotic body is unavailable, uneconomical, unreliable, or already occupied.
The term does not imply that a robot previously held the job and was replaced by a person. It refers to the direction of the production counterfactual. The company would prefer or eventually expect robotic execution, but human embodiment remains the available substitute.
4.1 Distinguishing reverse substitution from residual human work
Reverse robotic substitution differs from several familiar arrangements. In conventional residual work, automation performs most of a process while humans retain the unusually difficult portions. In human-in-the-loop systems, a person authorizes decisions, handles exceptions, or supervises an automated process. In teleoperation, a human remotely controls a robot.
A human occupying a robot slot may perform the main physical workflow rather than merely handling exceptions. The AI can already understand and coordinate the operation. The missing component is an affordable and available machine body. The person serves as the actuation layer for artificial cognition.
The resulting division of labor can be expressed as:
\text{AI planning and control} + \text{human mobility and manipulation} = \text{temporarily completed robot-native process}.
Historically, industrial machines supplied force and repetition while humans supplied judgment and adaptability. Reverse robotic substitution could assign judgment, sequencing, routing, and monitoring to AI while the human supplies reach, locomotion, dexterity, balance, and physical presence.
4.2 AI-directed embodiment labor
This arrangement creates AI-directed embodiment labor:
AI-directed embodiment labor is work in which an artificial system supplies most of the planning, procedural knowledge, task sequencing, and monitoring, while a person supplies the body required to act in the physical environment.
A worker might arrive at a facility with little knowledge of its internal layout. An AI system could direct the person to a station, identify the relevant object through augmented reality, show the correct grasp or orientation, verify completion through computer vision, and issue the next instruction.
The system can adapt its guidance to the worker. A novice may receive explicit step-by-step directions. An experienced worker may receive compressed task-level goals. Language, pace, interface modality, lifting limits, and route length can be personalized.
The person does not need to internalize the whole occupation. The production system remembers the inventory, sequencing rules, quality standards, exception procedures, and current state. This externalization of occupational knowledge could drastically reduce onboarding time.
4.3 The procedural exoskeleton
An AI guidance system of this kind can be understood as a procedural exoskeleton:
A procedural exoskeleton is an artificial system that supplements a worker’s memory, navigation, sequencing, decision-making, and task-specific knowledge in real time.
A physical exoskeleton supplements force. A procedural exoskeleton supplements competence. It enables a broadly capable person to enter a relatively unfamiliar environment and perform useful actions without first building an extensive internal model of the occupation.
Augmented-reality guidance already provides an early approximation. Research on warehouse order picking has found that AR smart glasses can improve speed and reduce errors, although effects on well-being and perceived productivity vary across workers and depend partly on technological familiarity. These results concern current support systems rather than full AI-generated embodiment labor, but they demonstrate that procedural knowledge can be delivered through wearable interfaces during physical work.
As perception and language models improve, the interface could move beyond fixed instructions. The AI may observe the environment, recognize that a box is damaged, revise the workflow, answer spoken questions, demonstrate an unfamiliar manipulation, and decide whether the worker should proceed or summon specialized assistance.
5. The Embodiment Gig Economy
The labor-market mechanism most compatible with reverse robotic substitution is an on-demand platform. A company facing temporary robot scarcity may not want to create permanent human careers inside roles scheduled for automation. A displaced worker may not want to retrain for a conventional multiyear occupation. Both parties can favor a short-duration exchange.
The embodiment gig economy can be defined as:
An on-demand labor market in which people provide temporary physical presence, movement, dexterity, sensory access, or force to AI-directed and robot-oriented production systems.
The platform matches physical tasks with locally available bodies. It may manage credentialing, scheduling, routing, instructions, performance verification, payment, insurance, and reputation.
5.1 Existing institutional precursors
On-demand physical labor platforms already operate in logistics, manufacturing, delivery, hospitality, and warehousing. Veryable, for example, describes a marketplace in which businesses post short-duration opportunities and qualified workers select or bid for them, with both parties rating one another afterward. Its model is designed to let companies expand labor capacity in response to fluctuating demand without maintaining the maximum required permanent workforce.
The platform model does not inherently require independent-contractor status. A natural experiment involving a parcel-delivery platform found that reclassifying workers as employees preserved scheduling flexibility while increasing hours and improving the match between scheduled and active work time. The result suggests that flexible, application-mediated labor can coexist with an employment relationship and associated protections.
Current platforms typically match people to recognizable human jobs or shifts. The embodiment gig economy would extend the platform deeper into the production process. It would not merely locate a warehouse worker. It would generate a stream of bounded physical actions, adapt those actions to the individual, and verify their execution.
5.2 From jobs to task streams
The unit of labor could shift from the occupation to the shift, and eventually from the shift to the task stream. A worker might open an application and select a four-hour assignment based on location, compensation, physical requirements, environmental conditions, and expected exertion.
After arrival, the person could receive a series of tasks from the workplace AI. The tasks may be related only by the fact that they require a mobile, dexterous body. The worker could stock a station, inspect a seal, move a flexible object, clear an obstruction for a robot, photograph an anomaly, and assemble an unfamiliar component under visual guidance.
This creates a Human Embodiment Task, abbreviated HET:
A Human Embodiment Task is a bounded physical action or action sequence assigned to a person because biological embodiment is available while suitable artificial embodiment is unavailable, fully utilized, or uneconomical.
The term parallels the Human Intelligence Tasks of digital crowdsourcing, but the scarce input has changed. The platform does not primarily need the worker’s knowledge. It needs the worker’s body.
A sufficiently integrated platform could function as an interface for requesting physical action. An AI system managing a warehouse might determine that no robot is available to remove a misplaced object. It requests a nearby qualified person, provides access instructions, guides the intervention, verifies completion, and issues payment. Human physical action becomes a callable service within a larger automated system.
5.3 The embodiment passport
Low barriers to entry do not eliminate the need for safety and qualification. The platform must know what each worker can perform. A conventional résumé is poorly suited to dynamically assigning physical microtasks.
An embodiment passport would provide a portable, machine-readable profile of verified physical capabilities and constraints. It might include lifting limits, mobility, balance, visual or auditory accommodations, equipment certifications, tolerance for particular environments, prior task success, ergonomic restrictions, language preferences, and availability.
The embodiment passport should not become a permanent biological ranking system controlled by employers. Workers would need authority over what information is collected, which attributes are disclosed, how assessments can be challenged, and whether data follow them across platforms. Health information requires especially strong protection.
Used appropriately, the passport allows task adaptation rather than exclusion. Two workers at the same facility may receive different loads, routes, shift lengths, break schedules, or instruction modalities. The economic system constructs a temporary role around the person’s safe capabilities.
5.4 Why drop-in embodiment work may become attractive
Conventional employment imposes large transaction costs on both sides. Applicants prepare résumés, attend interviews, present a narrative of long-term interest, complete onboarding, learn local procedures, and commit to recurring schedules. Employers invest in selection and training because the worker must retain substantial occupational knowledge.
AI-directed work can reduce these costs. A person may enter a facility, complete a short safety orientation, put on an instructional interface, and begin performing bounded tasks. The platform can retain the site-specific knowledge and update it continuously.
This creates drop-in embodiment labor, meaning short-duration physical work with limited site-specific preparation because AI supplies the necessary procedural support. Such work may appeal to people who need transitional income, are waiting for social-support institutions to change, are pursuing education or creative projects, have caregiving responsibilities, or simply reject the demand that every source of income become a professional identity.
The explicit impermanence may even be experienced as a form of honesty. A company does not ask the worker to express lifelong enthusiasm for a position scheduled for automation. The worker does not claim a long-term commitment. The company needs embodiment for a defined period, and the worker wants compensation for providing it.
5.5 Biological reserve capacity
Humans may remain useful even after substantial robot deployment. Demand spikes, breakdowns, recalls, disasters, unusual objects, remote locations, and novel edge cases can temporarily exceed machine capacity. A flexible human labor pool can serve as biological reserve capacity.
The platform may summon people whenever robotic utilization approaches its limit. Workers could cover seasonal peaks without requiring companies to purchase robots that remain idle during the rest of the year. They could also provide redundancy when a robotic fleet experiences a software failure or component shortage.
Biological reserve capacity is likely to decline over time, but it need not disappear immediately. Human bodies are already manufactured, widely distributed, self-repairing within limits, capable of operating in diverse environments, and supported by existing transportation and housing systems. These characteristics can temporarily offset their disadvantages in speed, consistency, and operating cost.
6. A Formal Model of Robot-Slot Labor
Let E(t) represent the quantity of embodied work demanded by robot-oriented production systems at time t, measured in standardized robot-equivalent task-hours. Let R(t) represent the embodied work capacity of the available robot fleet. Let c(t) represent the fraction of unfilled robot-oriented work that remains compatible with human execution, and let \rho_H(t) represent average human productivity relative to the standardized task unit.
The demand for human labor inside robot slots can be approximated as:
H_R(t) = \frac{c(t)\max [0,E(t)-R(t)]}{\rho_H(t)}.
This expression produces several important implications.
First, human robot-slot labor arises only where robot-oriented demand exceeds robotic capacity. A company may have a fully specified robotic workflow while lacking enough machines to execute it.
Second, anticipatory robotification can increase E(t) faster than robots are produced. New warehouses, factories, service processes, and logistics systems may be designed around machine labor. If the creation of robot slots outpaces growth in R(t), demand for human substitutes can rise.
Third, c(t) can initially remain high because early robotic workflows operate inside human environments or deliberately preserve a human mode. As robotic abundance increases, firms may cease paying the costs associated with human compatibility. Work speeds rise, safety separations disappear, environments become inhospitable, and specialized machine morphologies replace humanoids. The value of c(t) then declines.
These dynamics predict a hump-shaped trajectory:
\text{limited robotification} \rightarrow \text{rapid creation of robot slots} \rightarrow \text{human-filled capacity gap} \rightarrow \text{robotic saturation} \rightarrow \text{declining human compatibility}.
Human embodiment work may therefore expand during part of a transition whose endpoint contains far less human labor.
6.1 Demand expansion through artificial cognition
The quantity E(t) need not be fixed. Cheap artificial cognition may increase the volume of physical activity demanded. AI systems may design more products, personalize more services, identify more repairs, initiate more experiments, and coordinate more complex supply chains. Each digital decision can create a requirement for physical execution.
An AI scientific system can generate thousands of potential experiments, but laboratory capacity must perform them. An AI retail system can optimize individualized fulfillment, but products must still be moved. An AI construction system can design buildings rapidly, but materials must be assembled. Artificial cognition can consequently deepen the embodiment shortage by generating plans faster than machines can enact them.
This is one reason human robot-slot work may become substantial rather than merely residual. The growing cognitive system may create an enormous backlog of desired physical action while the robot fleet remains constrained.
6.2 Ambiguous wage effects
Robot scarcity places upward pressure on the value of human embodiment. When a production line is waiting for an unavailable machine, a person capable of keeping it operating can create substantial value. Urgent, dangerous, remote, unpleasant, or highly variable assignments may command scarcity premiums.
Cognitive displacement places downward pressure on wages. If large numbers of former office workers seek accessible physical shifts, the supply of available human bodies may exceed demand. Platforms can use ratings, dynamic pricing, and automated allocation to intensify competition among workers.
The wage outcome is therefore theoretically ambiguous. Some tasks may resemble surge-priced ride-hailing, with high compensation for immediate local availability. Others may resemble digital microtask markets, with many interchangeable workers competing for fragmented assignments. Labor law, bargaining institutions, platform concentration, and social insurance will strongly influence which tendency dominates.
7. From Knowledge Work to Embodiment Work
If cognitive automation advances faster than robotic capacity, labor-market displacement will not proceed through a simple division between employed manual workers and unemployed knowledge workers. Some people whose former expertise has lost market value will seek income in areas where human bodies remain useful.
This transition will be physically and psychologically uneven. A displaced administrator, programmer, researcher, designer, accountant, or legal worker may have little interest in establishing a second conventional career in warehousing or manufacturing. Drop-in embodiment labor offers a different proposition: perform a bounded physical shift without claiming a new occupational identity.
AI instruction can reduce the knowledge mismatch. It cannot eliminate differences in strength, disability, age, health, location, transportation, family obligations, or tolerance for physical conditions. High cognitive exposure and available physical jobs can therefore coexist with unemployment because people and tasks are not frictionlessly interchangeable.
The embodiment gig economy may nevertheless be attractive to a meaningful population. A person could work only when income is needed, choose assignments compatible with current physical condition, and preserve time for family, scholarship, art, community activity, or attempts to build a new livelihood. For some participants, physical work may offer movement and environmental variety after years of screen-based employment.
The distributional consequences would remain severe. Workers with chronic illness, impaired mobility, advanced age, or caregiving duties may be unable to participate. A social system that offers physical gig work as the only bridge between cognitive automation and income security would exclude many of the people most affected by technological displacement.
Robot-slot labor should therefore be understood as a possible transitional opportunity rather than a substitute for social provision. It may delay hardship for some people before universal basic income, social dividends, expanded public services, or other post-employment institutions emerge. It cannot provide a universal answer to technological unemployment.
8. Humans as Models for Robots
People occupying robot slots would generate unusually valuable training data. Cameras, wearable sensors, task logs, and augmented-reality interfaces could record how workers grasp irregular objects, recover from mistakes, navigate clutter, interpret ambiguous instructions, detect damage, and adapt when the modeled environment differs from reality.
Robot learning already depends heavily on demonstrations. The Open X-Embodiment project pooled more than one million trajectories across 22 robot embodiments and found that training across diverse platforms improved transfer and generalization. More recent efforts have increasingly explored learning from human video and cross-embodiment data.
Figure has reported training a humanoid navigation system using egocentric human video and has described a larger effort to collect human goal-directed behavior across homes, offices, and logistics environments. Company claims require independent evaluation, but the developmental strategy demonstrates the economic value assigned to ordinary human movement data.
A human robot-slot worker may therefore perform two forms of labor simultaneously. The first is direct production: moving the component, stocking the shelf, inspecting the package, or clearing the obstruction. The second is data production: demonstrating how an adaptable body solves the task.
This dual use creates an asymmetry. The company receives immediate output and training information that may reduce its future dependence on the worker. The person is paid for the present shift while contributing to the elimination of future shifts.
A fair labor framework should recognize robot-training compensation:
Robot-training compensation is additional payment or continuing economic participation for workers whose movements, corrections, decisions, or edge-case solutions are used to train robotic systems.
Compensation could take several forms. Workers might receive a data premium, a licensing payment, a collective automation fund contribution, or a share of savings produced by models trained on their demonstrations. Collective mechanisms may be preferable to individualized royalties because robotic policies are trained on aggregated data and the causal contribution of one demonstration can be difficult to measure.
Workers also need the right to know whether their behavior is being recorded for training. Consent to workplace monitoring should not automatically imply consent to unrestricted model development. Production data, safety monitoring, biometric inference, and robot-training data should be treated as distinct categories.
9. Robotic Normativity and the Experience of Work
Anticipatory robotification changes more than who performs a task. It changes the standard by which the task is understood.
Robotic normativity refers to the use of machine performance as the reference model for human work. A robot’s path, timing, consistency, and movement economy become the ideal procedure. People are instructed to approximate that procedure while they remain in the slot.
The worker may watch a robot perform the task, follow an animated overlay derived from a robotic policy, or receive corrections whenever their movements diverge from the reference sequence. Artificial intelligence can compare observed human behavior against the model and identify unnecessary steps, delays, or orientation errors.
Some machine-derived procedures may improve work. Optimized routing can reduce walking, automated checking can prevent mistakes, and better presentation of parts can reduce awkward reaches. The robot’s performance may reveal unnecessary inefficiencies inherited from older production methods.
Other procedures can be hazardous when transferred directly to humans. A robot does not experience repetitive strain, humiliation, boredom, attentional fatigue, or anxiety in the ordinary biological sense. It may operate at a pace that is safe for actuators but unsafe for muscles, tendons, joints, and attention. Human mode must therefore optimize for health rather than merely slowing the robot’s trajectory.
9.1 Algorithmic control
Algorithmic management already instructs, monitors, and evaluates workers. An OECD survey of more than 6,000 managers across six countries found broad adoption of at least one algorithmic management tool, including systems that allocate activities, provide task instructions, track work speed, monitor location, set targets, and sanction or reward performance. Managers perceived benefits to decision quality, while concerns included accountability, explainability, and inadequate protection of physical and mental health.
The embodiment gig economy could intensify this form of control because the platform would mediate nearly every aspect of work. It might determine who gains access to a shift, generate the task sequence, set the pace, track the body, evaluate completion, calculate pay, and update the worker’s future eligibility.
An OECD experiment examining algorithmic management in simulated manufacturing found that workers and managers were particularly sensitive to system strictness and worker autonomy. Participants objected to systems that imposed inflexible task order or delivered intrusive alerts, while a design in which the system helped an inexperienced worker learn a task was broadly acceptable. These findings suggest that the same technology can function either as a supportive procedural exoskeleton or as an invasive mechanism of behavioral control.
9.2 The autonomy paradox
Drop-in embodiment labor can increase autonomy at the level of scheduling while reducing autonomy inside the task. A person may freely decide whether to work on Tuesday afternoon, yet experience almost no discretion after entering the facility.
This produces an autonomy paradox. The platform offers freedom from long-term employment commitments while imposing granular control over movement, timing, and procedure. Whether workers experience the arrangement as liberating or oppressive will depend on compensation, voluntariness, task design, availability of alternatives, and the degree to which the system permits judgment and refusal.
A humane implementation would allow workers to pause, question an instruction, report that the modeled environment is wrong, reject an unsafe action, and explain a more efficient procedure. The system should learn from worker expertise rather than treating every deviation as failure.
10. Developmental Stages
The framework suggests a sequence of overlapping stages rather than a single transition date.
10.1 Cognitive acceleration
Artificial systems automate increasing portions of planning, communication, analysis, software use, and coordination. Knowledge work is reorganized, and some workers experience reduced demand for their previous skills. Physical tasks remain comparatively protected by the cost and difficulty of embodiment.
10.2 Anticipatory design
Companies identify processes that will eventually be performed by robots. They standardize objects, digitize workflows, install machine-readable infrastructure, develop task graphs, and create robot-compatible facilities. Humans continue performing much of the work, but the production model increasingly treats them as temporary occupants.
10.3 Reverse substitution
Robot slots proliferate faster than robotic bodies. Companies use employees, temporary staff, and gig workers to fill the difference. AI systems provide real-time guidance and supervision, reducing the site-specific knowledge needed from each person.
10.4 Embodiment-platform expansion
On-demand platforms become integrated with production-management systems. Workers maintain embodiment passports, accept short shifts, and receive task streams through wearable devices. Human Embodiment Tasks become a recognizable category of economic activity.
10.5 Substrate-flexible production
Facilities can interchange humans and robots within a common task architecture. Humans provide peak capacity, exception handling, unusual manipulation, and geographic coverage. Robots occupy the most repetitive and predictable positions.
10.6 Robotic saturation
Robot manufacturing, maintenance, and deployment approach the scale of robot-oriented demand. Robotic performance improves, operating costs fall, and machine capacity becomes available for most routine physical positions. Demand for human substitutes declines.
10.7 Human-mode withdrawal
Companies redesign facilities without preserving biological compatibility. Speeds increase, specialized morphologies proliferate, and infrastructure no longer assumes human access. People cease to be merely less efficient occupants of robot slots and become physically inadmissible to many production environments.
11. Testable Predictions
11.1 Robot-oriented workflows will precede robot availability
Companies preparing for robotic deployment should begin standardizing tasks and facilities before acquiring enough robots. Evidence would include workplaces in which people follow machine-readable task graphs, robot-compatible object presentation, and AI-generated procedures while robotic fleet utilization remains low.
11.2 Onboarding time will decline in instrumented physical work
AI guidance and augmented-reality instruction should reduce the amount of site-specific knowledge workers must memorize. The effect should be strongest in structured facilities with digitized inventories, good sensor coverage, and bounded task variation.
11.3 The same task representation will be executed by different embodiments
Substrate-flexible systems should represent goals independently of the executor. Researchers should observe common task graphs being translated into distinct robot policies and human-safe instructions.
11.4 Human demand will rise temporarily in rapidly robotifying sectors
Where robot-oriented work expands faster than robotic capacity, demand for temporary physical labor should initially rise. This increase should reverse once local robot supply, uptime, and productivity cross the required threshold.
11.5 Physical gig platforms will recruit displaced cognitive workers
If knowledge-work displacement precedes robot saturation, platform data should show increasing participation by people with prior professional, administrative, creative, technical, or clerical employment. Their participation may be episodic rather than a permanent occupational transition.
11.6 Novelty will predict continued human occupation
Humans should remain longer in robot slots involving environmental variability, ambiguous object states, unusual exceptions, or sparse training data. Predictable, high-volume tasks should transition to robots earlier.
11.7 Human work data will accelerate robotic substitution
Instrumented workplaces should use worker demonstrations, corrections, and exception handling to train robotic systems. Positions with extensive data collection should automate faster than comparable positions where worker behavior is not captured.
11.8 Human compatibility will eventually decline
Early robot-oriented facilities should retain human walkways, safety controls, lighting, climate regulation, and accessible interfaces. Mature robot-native facilities should reduce these accommodations when their economic value falls.
11.9 Wages will become more geographically and temporally volatile
Local robot shortages, breakdowns, and demand spikes should produce temporary wage premiums. Regions with large pools of displaced workers and few alternative jobs may show the opposite pattern, with low compensation despite high technological sophistication.
11.10 Algorithmic-management design will predict retention and safety
Systems that allow worker feedback, discretion, and refusal should produce better safety and retention than systems that impose rigid machine-derived timing. Workplace outcomes should vary according to the design of human mode rather than according to AI adoption alone.
12. Policy Implications
12.1 Require a genuine human mode
Any robot-oriented workplace that admits human workers should maintain a validated human execution policy. Safety standards should address repetition, lifting, heat, noise, attentional load, walking distance, recovery time, and the interaction between people and high-speed machines.
Human mode should be independently auditable. Companies should not be permitted to classify a machine procedure as human-safe merely because a person can physically complete it once.
12.2 Preserve the right to refuse unsafe instructions
Workers must be able to pause or reject an AI-generated task without automatic punishment when conditions differ from the model or when the action appears unsafe. A rapid human appeal channel should exist for account suspensions, performance penalties, and disputed safety events.
The system should distinguish disobedience from useful correction. Workers often possess local knowledge unavailable to the central model, especially during unusual events.
12.3 Separate flexibility from precarity
Application-mediated scheduling does not require the removal of employment protections. Minimum compensation, injury coverage, unemployment insurance, paid waiting time, and collective representation can coexist with flexible shifts. The evidence from platform reclassification suggests that employee status need not eliminate scheduling flexibility.
Compensation should account for travel, mandatory preparation, equipment fitting, security delays, and time spent waiting for task assignment. Otherwise, nominally attractive hourly rates can conceal substantial unpaid labor.
12.4 Establish worker control over embodiment data
Motion traces, gaze, voice, physiological signals, error patterns, fatigue estimates, and health inferences can reveal sensitive information. Workers should know which data are collected, how long they are retained, whether they train robots, and whether they are shared across employers.
An embodiment passport should be portable and worker-controlled. Platforms should not be able to create permanent, opaque biological risk scores that exclude people from work without explanation or appeal.
12.5 Compensate robot-training labor
When worker behavior becomes training data, compensation should extend beyond ordinary task pay. Collective automation funds may be administratively simpler than tracing the contribution of each individual movement to a later model.
A portion of productivity gains could be directed toward the workers and communities that supplied the demonstrations. Such arrangements would recognize that the robotic system was trained through accumulated human practical intelligence, even when the final machine no longer resembles its teachers.
12.6 Prepare for the end of the bridge
Robot-slot labor is inherently transitional. Governments should not interpret temporary growth in physical gig work as evidence that technological unemployment has been solved. The same companies creating these positions may be actively developing the machines that eliminate them.
Income guarantees, public services, shorter working weeks, social wealth funds, ownership participation, and automation dividends should be designed before robot saturation. Waiting for embodiment work to disappear would recreate the institutional lag that made the bridge necessary.
13. Qualifications and Alternative Trajectories
Several conditions could weaken or prevent the proposed sequence. Cognitive automation may progress more slowly than expected, leaving knowledge work broadly human-dependent while robots gradually diffuse through physical industries. In that case, there would be less displaced labor available for the embodiment gig economy.
Robot production may scale faster than assumed. Standardized actuators, modular designs, automated factories, and robots building robots could compress the embodiment lag. Apptronik and Jabil have already described plans to use newly manufactured Apollo units in the facilities that produce them, creating a potential manufacturing feedback loop.
Organizations may prefer specialized automation over humanoids. Conveyor systems, autonomous mobile robots, fixed arms, and redesigned products can outperform general-purpose bodies in controlled settings. The relevant scarcity would then involve integrated robotic systems rather than humanoid counts.
Human labor may remain cheaper than robotics in many regions. Even where a robot is technically superior, financing costs, maintenance, energy, downtime, liability, and integration can preserve human employment. The robot may become the conceptual ideal without becoming the immediate economic choice.
Consumers and regulators may also preserve human roles. Care, hospitality, education, public authority, and personal services can involve social preferences that are not reducible to technical task performance. A robot-oriented production system may remain politically or culturally unacceptable even after it becomes technically feasible.
Finally, universal basic income or another income institution could precede large-scale reverse robotic substitution. People might then choose embodiment shifts for supplemental income, structure, exercise, or social contact rather than economic necessity. The embodiment gig economy could remain popular under such conditions, but its bargaining dynamics would be substantially more favorable to workers.
14. Conclusion
Artificial intelligence may automate the direction of work before industry can automate its physical execution. During the resulting embodiment lag, firms will possess abundant artificial planning, reasoning, coordination, and procedural knowledge while lacking enough machine bodies to carry those plans into the world.
Anticipatory robotification describes the reorganization that follows. Companies begin designing jobs and facilities for robots before robotic fleets can occupy them. Human workers remain inside the system, but the robot becomes the normative worker, and the person becomes a temporary physical substitute.
Reverse robotic substitution captures this inversion. People fill robot slots because bodies are scarce. Artificial intelligence tells them where to go, what to manipulate, how to perform the action, and whether the result is acceptable. The human supplies mobile, dexterous embodiment to a process whose cognition increasingly resides elsewhere.
The embodiment gig economy could make this work widely accessible. Applications can match people to local physical demand, procedural exoskeletons can lower training barriers, and embodiment passports can allow task allocation according to verified capabilities. Former knowledge workers may enter warehouses, factories, retail environments, laboratories, and service sites without treating the assignment as a new lifelong career.
Human demand may temporarily increase as robot-oriented production expands faster than robot manufacturing. This growth would contain its own endpoint. Every instrumented shift can generate training data, every standardized task can become easier to automate, and every new robot can reduce the remaining capacity gap. The bridge is built while the far side is advancing toward it.
The most striking feature of the transition is therefore its reversal of ordinary technological substitution. For a limited historical interval, humans may be employed because they are the most abundant available robots. They may watch machines to learn how their work should be performed, receive instructions from artificial intelligence, and serve as biological reserve capacity for a production system waiting for sufficient hardware.
Human labor would persist inside an economy that had already selected its successor. Robotification would arrive before the robots.
References
Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30.
Apptronik. (2024). Apptronik and Mercedes-Benz enter commercial agreement.
Apptronik. (2025). Apptronik and Jabil collaborate to scale production of Apollo humanoid robots and deploy in manufacturing operations.
Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30.
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.
Figure AI. (2025). F.02 contributed to the production of 30,000 cars at BMW.
Figure AI. (2025). Project Go-Big: Internet-scale humanoid pretraining and direct human-to-robot transfer.
Figure AI. (2026). Introducing Helix 02: Full-body autonomy.
Figure AI. (2026). Ramping Figure 03 production.
Gmyrek, P., Berg, J., Kaminski, K., Konopczyński, F., Ładna, A., Nafradi, B., Rosłaniec, K., & Troszyński, M. (2025). Generative AI and jobs: A refined global index of occupational exposure. ILO Working Paper 140.
International Energy Agency. (2026). Rare earth elements.
International Federation of Robotics. (2025). World Robotics 2025: Industrial Robots.
International Federation of Robotics. (2025). World Robotics 2025: Service Robots.
Johnston, H., Ergun, O., Schor, J., & Chen, L. (2024). Employment status and the on-demand economy: A natural experiment on reclassification. Socio-Economic Review, 22(1), 169–194.
Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366–410.
Milanez, A., Lemmens, A., & Ruggiu, C. (2025). Algorithmic management in the workplace: New evidence from an OECD employer survey. OECD Artificial Intelligence Papers, No. 31.
OECD. (2025). Exploring win-win outcomes of algorithmic management.
Open X-Embodiment Collaboration. (2023). Open X-Embodiment: Robotic learning datasets and RT-X models. arXiv:2310.08864.
Vallas, S., & Schor, J. B. (2020). What do platforms do? Understanding the gig economy. Annual Review of Sociology, 46, 273–294.
Windhausen, A., Heller, J., Hilken, T., Mahr, D., Di Palma, R., & Quintens, L. (2024). Exploring the impact of augmented reality smart glasses on worker well-being in warehouse order picking. Computers in Human Behavior, 155, 108153
Leave a Reply