Humanoids are being deployed on assembly lines, while robots currently documented in food service are mostly service or delivery robots, not necessarily humanoid. According to RCSV, trained teleoperators in India, the Philippines, and Eastern Europe are paid between 22 and 55 dollars per hour for data collection and teleoperation services. American specialists capture the other end of the chain, at 65–120 dollars per hour. Between the two, a global labor market is organizing according to a logic that digital platforms have already traced: value rises, costs fall, and geography does the rest.
The essentials
- Humanoid robotics is creating a layer of teleoperation and annotation whose training costs have fallen 40 to 60% between 2024 and 2026 according to RCSV.
- Teleoperation operators in Asia and Eastern Europe earn between 22 and 55 $/h; American specialists in the same chain earn 65 to 120 (Robotics Center AI, State of Robotics 2026).
- RCSV indicates that more than 340 fast-food restaurant sites use a robot and that a robot can break even in 3–4 years when labor costs exceed 18 $/h.
- Teleoperation can geographically separate operators and robots, but the global distribution of value created is not established by this source.
- The central question is whether this layer is a stepping stone toward robot autonomy or a devalued job that perpetuates itself, and signals through 2030 will separate the two trajectories.
Pizza and a click 12,000 kilometers away
RCSV documents teleoperation and data collection services, but not an established model of humanoid fast-food robots piloted from Asia or Eastern Europe. The State of Robotics 2026 from Robotics Center AI lists more than 340 robot deployments in fast-food service. According to RCSV, a cooking or frying robot would break even in three to four years at labor costs above 18 $/h; the report does not quantify the 15 $/h case. RCSV advances a breakeven estimate at high labor costs, without documented conditions of low-cost supervision.
This architecture is not a childhood defect. It is the product of precise calculation. The 40 to 60% reduction in annotation work is attributed by RCSV to its annotation pipelines. The fall in teleoperation costs facilitates pilots, while RCSV links the breakeven of certain kitchen robots to labor costs above 18 $/h. The circle is virtuous for deployers, far less so for data producers.
This mechanism aligns with the analysis published on emerging countries facing polarization by robots: technology redistributes work geographically, and low-cost countries generally end up in segments where the value created is lowest.
Two labor markets, one production chain
The State of Robotics 2026 report documents two pay levels that barely ever intersect. Teleoperation operators and data annotators in Asia and Eastern Europe earn between 22 and 55 dollars per hour. American engineers and specialists who design systems, validate models, and oversee deployments earn between 65 and 120 dollars per hour.
The gap is structural as much as it is about wages. The first group produces raw material: training data, trajectory corrections, supervision signals. Ownership of models and data depends on contracts and companies. The second group has access to a significant share of the rent from these models once trained. The frontier between the two groups is geographical as much as it is professional.
This pattern resembles what the International Labour Organization (ILO) documents in its Digital Labour Platforms 2025 report on digital work platforms: a global division of labor where value production and value capture are systematically separated, the former localized in low-cost economies, the latter in economies with high technological capital. This pattern applies to humanoid robotics, with one difference: demand for teleoperators could increase if deployments accelerate. What makes it potentially more dangerous if models reach autonomy before workers have had a chance to upskill.
The $18 threshold reveals deployment logic
The 340 documented deployments are concentrated in the United States, Japan, and South Korea. RCSV advances a breakeven estimate for certain kitchen robots when labor costs exceed 18 $/h, without establishing that all sites exceed this threshold. Below it, the calculation reverses.
This threshold has a direct consequence for the geography of work. In 2024, Asia represented 74% of global industrial robot installations and China 54%. They supply a significant share of the cognitive raw material that makes deployments elsewhere profitable. They export value without hosting the machines that this value makes possible to finance.
The analogy with physical raw materials is tempting. Africa produces 70% of the world’s cobalt but lets the added value slip away: humanoid robotics could install similar mechanics for data. Producing countries supply the resource, technologically dominant countries refine it and capture the rent.
The difference from cobalt is that data loses value quickly. An annotator trains a model that, once trained, no longer needs them. The cobalt miner remains indispensable as long as the mine produces. Cognitive resources have a lifespan that physical resources do not.
The fall in annotation costs and its consequences
A 40 to 60% reduction in annotation costs over two years reflects a market trajectory, beyond simple efficiency gains. Automated annotation tools are improving, models are learning to self-correct on certain tasks, and demand for human work per unit of data produced is declining.
For teleoperators and annotators, this signal is ambiguous. If global robot demand grows faster than annotation tool productivity, the volume of available work can increase even if unit cost falls. If automated annotation tools absorb the most repetitive tasks, remaining human tasks are more complex, but not necessarily better paid if the supply of skilled workers remains abundant.
RCSV notes that the reduction in robot production costs follows a curve similar to that of training data. The two curves, robot cost, data cost, fall together, which accelerates the breakeven point in new markets. For deployers, this is an opportunity. For data producers, it is pressure on hourly rates as the supply of workers skilled in these tasks increases globally.
Teleoperation in 2035: transition employment or permanent segment
The permanence of teleoperation is a possible factor, among several technological, economic, and regulatory determinants of the labor market. Both scenarios have internally coherent logics.
One scenario envisions that robots progressively reach a level of operational autonomy sufficient to reduce real-time human supervision on standardized tasks. Fast food, simple assembly, repetitive logistics: these areas disappear from the teleoperators’ remit. Demand for supervision collapses before workers concerned have had time to retrain. Low-cost countries then lose both the transitory rent and the upskilling opportunity it might have funded.
Another scenario envisions that robots maintain a residual error rate that could require prolonged human supervision. Non-standardized tasks, unpredictable environments, complex interactions, and new situations remain beyond the reach of complete autonomy. Teleoperation then stabilizes as a permanent segment, with growing volumes as deployments accelerate. The issue becomes compensation: will the current 22 to 55 $/h remain the norm, or will they converge toward Western levels if demand exceeds available supply.
Both scenarios share a common need: traceability of value produced by data workers. Currently, annotators in Asia and Eastern Europe rarely have established rights over the value created by models they helped train. The source documents hourly rates, without establishing the full contractual or economic rights of workers. If robots reach autonomy quickly, this relationship will have been the only one, and its brevity will have left no trace in the annotator’s future skills or income.
Two signals will settle the trajectories by 2030. The first is the evolution of median hourly rates for teleoperators in Asia and Eastern Europe: stagnation of hourly rates would be a possible index, but insufficient on its own to conclude permanent devaluation; convergence toward Western levels could be compatible with supply tension without allowing it to be established on its own. The second is the share of fast-food deployments operating without real-time human supervision above the 18 $/h threshold: if this share increases significantly before 2030, the first scenario gains in probability.
Errors already made by platforms
Digital platforms of the 2010s followed a similar path. Uber, Amazon Mechanical Turk, content platforms: all created fragmented labor markets, geographically dispersed, where value producers captured a minority fraction of the rent. None spontaneously established the redistribution mechanisms that ILO researchers were recommending at the time.
Humanoid robotics arrives with a potential difference: it is physical, therefore visible, therefore more easily regulated than digital platforms. A robot in a Los Angeles fast-food restaurant has an identifiable deployer, a localized manufacturer, a traceable value chain. The data it consumed to learn to function was produced somewhere, by someone, at a documented price.
This traceability makes possible mechanisms that were impossible to apply to purely digital platforms. A levy on training data collected abroad, redistributed as conversion funds to producing countries, would follow the logic of licensing rights applied to natural resources. Portable skills certifications for teleoperators, recognized independently of the client, would secure the professional trajectory in both autonomy scenarios. These paths exist in academic discussions and in some ILO reports; they have not yet found a political champion at the international level.
The architecture of a labor market freezes quickly: wage norms, statuses, and contractual practices established in the first years of a new sector tend to persist. The window for setting different rules in humanoid robotics is open, but it will not remain so indefinitely. As the analysis of German development shows, robots can create skilled jobs, but this result depends on deliberate institutional choices, not on technology alone.
Sources
- Robotics Center AI, State of Robotics 2026, https://www.roboticscenter.ai/en/state-of-robotics-2026
- Goldman Sachs, Humanoid Manufacturing Costs Analysis 2026, internal report cited in specialized press; not available in open access
- International Labour Organization (ILO), Digital Labour Platforms 2025, available at ilo.org
