There is a photograph of the humanoid robotics boom that nobody puts on a keynote slide: a factory worker in Texas, wearing a motion-capture suit, performing the same pick-and-place motion for hours so that a machine can one day do it without her. This month that image finally became a story. Reporting from The Information, relayed by Ars Technica, describes Tesla’s heavy reliance on imitation learning prompting the company to have factory workers in Texas and California wear special suits to record their physical movements while working. The workers, according to the report, complained, because they “knew the robots were designed to eventually replace them.”
Tesla’s response was organizational, not technological: it shifted data collection to dedicated teams and set up stand-alone training hubs. It is a small management decision inside a very large industrial bet, and it deserves more attention than it has received, because it names the contradiction at the center of the entire humanoid category in 2026. The robots being sold as autonomous labor run on human labor. The question is not whether that human layer exists. It is who pays for it, what it costs, and whether the industry’s economics survive its eventual thinning.
The scale behind the anecdote
The context makes the anecdote more than a human-interest detail. Tesla is now building several hundred Optimus units per week at Fremont, roughly ten times its Q2 output, with managers targeting a continuous automated line exceeding 1,000 robots per week by year-end and an eventual goal of 20,000 per week, Electrek reports. The line occupies the space where the Model S and Model X used to be built; Tesla ended both flagship vehicles in May and pulled staff onto the robot program.
Where the robots go is the more revealing number. Most units are used internally for testing, training and data collection. The ones working inside Tesla’s own factories are confined to tightly controlled, supervised areas and programmed for specific tasks rather than operating as general-purpose machines. The V3 robots currently in production are not even the commercial version, which still has to pass stricter durability thresholds. This is a fleet whose primary output, for now, is training data.
That is why the workers in mocap suits matter strategically and not just morally. Tesla has accumulated more than 500,000 hours of training data and wants to double that figure by year-end. It has moved much of its self-driving data annotation team onto Optimus, hired dedicated data collectors wearing camera helmets and motion-capture suits, and opened training hubs in Colorado, Arizona and Florida. The information-age comparison writes itself: an automaker that once annotated lane changes now pays humans to perform factory motions in costume, because the next model in the pipeline is a body rather than a car.
Why imitation learning demands humans
The dependency is not a Tesla quirk. It is a structural property of the current technical stack. Vision-language-action models, the family that includes Physical Intelligence’s π0, Figure’s Helix and NVIDIA’s GR00T, learn manipulation primarily from demonstration: recorded trajectories of humans solving tasks, mapped onto robot bodies. Simulation helps, and synthetic data increasingly fills gaps, but the sim-to-real transfer for contact-rich dexterous work remains imperfect enough that every serious program still buys, rents or employs human demonstrators.
The cost shows up in an uncomfortable metric. According to The Information’s sources, it still takes Optimus several days to learn even basic tasks, and the system behaves unpredictably in situations it has not been trained on. Tesla’s strategy is to build a library of primitive movements the robot can recombine for new jobs, which is a bet that task-level learning curves eventually flatten into skill composition. Until that bet pays out, every new task has a human data bill attached: operator hours, annotation hours, evaluation hours.
That bill is now a visible employment category. Aggregator sites such as robottraining.jobs track a live feed of teleoperation roles, demonstration-data gigs and embodied-AI evaluation work, alongside mass “record-and-earn” programs from Figure Index, DoorDash Tasks and micro1 that pay people to film themselves performing everyday household tasks. In September, Toborlife AI launched a full-body teleoperation system for the Unitree G1, packaging motion capture, training-data collection and dexterous control as a single workflow. A tools market for extracting human motion is forming in real time, which tells you what the demand side looks like.
The consumer layer: Expert Mode and the remote operator
The same architecture has quietly entered the home. 1X’s NEO, the humanoid most explicitly positioned for domestic deployment at $20,000 outright or $499 per month, is trained and supported through teleoperation: human operators remotely guide the robot through real homes, generating the demonstration data that future autonomy will be built on. The company’s Expert Mode extends the pattern into the product itself, allowing a human supervisor to take over when the robot encounters tasks it does not yet know, as Forbes detailed in a piece on consumer trust.
Forbes’ reporting on Tau, a cleaning-robot maker, makes the privacy posture explicit: its robots combine AI with a remote human operator who sees the home through the robot’s cameras, with microphones disabled and bedrooms and bathrooms excluded by default. Credit where due, those defaults are the right defaults. But notice what they concede. A robot marketed as autonomous is, in its early commercial life, a mobile camera with a stranger attached, bounded by policy rather than by capability. The industry’s honest disclosure problem is not hypothetical; it is a live consumer-trust question being negotiated one privacy setting at a time.
The commercial logic is coherent, even if the marketing sometimes is not. Teleoperation is how you deploy into environments you cannot control, collect the exact data those environments generate, and keep the customer relationship alive while the models improve. It is the FSD playbook transposed to hardware: ship the machine, harvest the fleet, promise the software will catch up. Tesla’s reported plan to lease rather than sell Optimus to a short list of companies whose facilities resemble its own factories fits the same template, with the added feature that leasing keeps fleet data rights, and the human labor embedded in it, entirely on Tesla’s side of the contract.
What the human layer does to the numbers
Yole Group’s 2026 humanoid market analysis counts more than 60 active manufacturers worldwide, over half of them in China and 12 in the United States, with cumulative category funding of roughly $9.8 billion since 2017. Against that capital base, the human training layer functions as a real operating cost that rarely appears in unit-economics decks: demonstration hours per task, teleoperator ratios per deployed fleet, annotation throughput, evaluation cycles. A humanoid that needs several days of supervised learning per task and a remote human on standby is not yet displacing labor. It is redistributing it, from the factory floor to the training hub, often at lower wages and with no claim on the asset being built.
The valuation environment is starting to ask exactly these questions. Even as deployments multiply, Benzinga notes that private-market humanoid valuations are drawing skepticism from investors including Vinod Khosla, whose firm backed some of the category’s foundational bets. Musk’s own framing, that Optimus could be “the biggest product ever,” now coexists with a documented history on this exact class of promise: in January 2025 he said Tesla would build roughly 10,000 Optimus robots that year with “several thousand” doing useful work by year-end, and by January 2026 he conceded that zero were doing useful work. The hardware, to Tesla’s credit, now genuinely scales. Fremont is producing several hundred robots a week with a path to 1,000. The unsolved variable is the same one that has always dogged the company’s autonomy claims, and it is a software variable: generalization, in robots that today still need days of training per basic task.
The reliability picture reinforces the point. The Information describes hands with more than 100 screws and small components still assembled by hand, fixture misalignment forcing rework, touch sensors unreliable enough that Tesla developed a replaceable “sensing glove” so the entire hand need not be swapped, and suppliers, many in China, that hold tolerance in prototype volumes but struggle at scale. Meanwhile XPeng has started running an automated humanoid production line in Guangzhou targeting 2027 commercial sales, Hyundai plans to deploy up to 25,000 Boston Dynamics Atlas robots, and Toyota is committing billions to factory robot upgrades. The race is crowded enough that hidden costs, and hidden labor, will differentiate balance sheets.
The ethics of training your replacement
There is a version of this story in which the mocap-suit worker is simply the latest entry in a long history of automation absorbing human skill without compensating its source. loom operators, switchboard operators, typesetters: nobody paid them royalties on the processes their work rationalized. The humanoid version has a sharper edge, because the worker is not adjacent to the substitution asset. The worker is the substitution asset. Every hour of recorded motion is a transfer of embodied expertise from a person who will be paid once for the hour to a machine that can replay it indefinitely.
The available policy responses are not exotic. Wage premiums or equity grants for data-collection roles would treat demonstration work as what it is: foundational intellectual property. Disclosure obligations would tell customers when a “robot” service includes remote humans, the same way call-center disclosure evolved. And deployment agreements, particularly the lease structures Tesla favors, could be required to specify data provenance, including whether the training corpus was collected from the workforce of the very facilities being automated. None of this is anti-robot. It is the difference between a labor transition and a labor extraction.
The unions have already located the fault line. Hyundai’s humanoid deployment plan surfaced precisely amid talks with striking workers; Toyota frames its 400,000-robot factory program around augmentation. Companies that want smooth adoption at scale have a direct incentive to make the training workforce whole, because the alternative is a workforce that declines to train, exactly as Tesla’s employees did, forcing the company to build parallel data-collection organizations at additional cost. Fremont’s training hubs are, among other things, the price of not having answered that question earlier.
When the humans thin out
None of this argues the human layer is permanent. The trajectory of the field is unmistakable: better sim-to-real, synthetic data generation, world models, and VLA architectures that compose skills instead of memorizing trajectories. The metric worth tracking is hours of human demonstration per new economically useful task. When that number falls from days toward hours, and then toward zero for whole task families, the economics of the category invert, and the current arrangements, the hubs, the teleoperators, the Expert Mode safety nets, become transitional scaffolding.
But that day is not today, and every serious operator knows it. Today the humanoid industry is a hybrid workforce in which silicon does the repetition and carbon does the learning. The honest formulation of the 2026 state of play is this: autonomous humanoids are being built by humans, wearing suits in Arizona, guiding arms through VR headsets, watching kitchens through remote cameras, and filming themselves folding laundry for piece rates. The companies that acknowledge this layer, price it, and pay for it will build better robots faster than the ones that pretend it isn’t there. The labor was always going to be part of the machine. The only question was whether it would be visible, and now it is.