Before a humanoid can do a task alone, a person must show it how, often thousands of times. And when it fails, another person quietly takes the controls.
A humanoid robot loads a dishwasher. The video is smooth and the caption says “autonomous”. What the viewer cannot see is the question that matters most in robotics today: was anyone else involved?
Very often the answer is yes. The current generation of humanoids depends on two kinds of human labour that rarely appear on screen. One group produces the data from which robots learn. The other steps in, remotely, when a robot cannot cope. Together they form a new workforce that the industry seldom mentions.
Teaching by repetition
Language models learned from text that already existed on the internet. Robots have no such library. There is no vast archive of recorded human movement, with the forces and adjustments involved in gripping a cup or turning a handle. It has to be made.
In early 2026 MIT Technology Review described how it is made. In Shanghai a worker spent a week wearing a virtual-reality headset and an exoskeleton, opening and closing a microwave door hundreds of times a day so that a robot could learn the motion. Elsewhere, employees of a delivery company wore movement sensors while shifting boxes, producing data for the machines that might one day shift them instead.
Homes are becoming data sources too. The American firm Figure announced a partnership with Brookfield, an investment company that manages about 100,000 residential units, to gather recordings of real household environments.
Each of these efforts turns ordinary activity into raw material. The worker is paid for the hours. The company keeps the recording, which can be reused indefinitely to train any number of machines.
The same logic explains an odd job advertised in San Francisco in 2026. A startup named Dimensional offered $20 an hour to people willing to walk a robot dog around the city, steering it with a remote control. The machine’s sensors recorded each walk. The company’s founder described the recordings as “ground truth” for teaching software to navigate, and admitted that AI agents are still poor at controlling hardware.
The operator in another room
The second kind of work is teleoperation: a person watches through the robot’s cameras and moves its body from a distance.
The clearest example is Neo, a home robot from the company 1X, sold for $20,000 or rented for $499 a month. Its factory in California began full production in April 2026. The company says Neo can fold towels, fetch objects and open doors. Early hands-on tests, however, found that it could not complete a single household task without a remote human operator.
1X does not hide this entirely. It calls the staff who take control “operators” and has built in safeguards: a light on the robot glows when a human is driving it, faces in the video feed are blurred, and owners can mark rooms that the robot and its operators may not enter. Even so, the arrangement means that a stranger may be looking into a customer’s home. The AI researcher Gary Marcus has predicted that the privacy concerns will be huge.
Teleoperation is not confined to homes. The same startup that paid dog walkers has described a security-patrol service in which a remote human stays in the loop.
Why companies stay quiet
Robotics firms have reasons to say little about this labour. Investors and customers pay for autonomy. A video in which a robot works alone is worth more than one in which it is visibly a puppet. So the human contribution tends to be mentioned in small print or not at all.
MIT Technology Review’s argument is that this silence misleads the public. People overestimate what robots can do by themselves, just as they once overestimated driver-assistance systems in cars. The article draws a parallel with the earlier history of AI, whose training data was labelled by large numbers of low-paid workers, often in poorer countries, who were also kept out of view.
Work that can move anywhere
There is a larger consequence. If a robot’s body can be in one country and its operator in another, physical work becomes exportable for the first time. A person in a low-wage country could, in principle, clean a hotel room or stack a shelf in a high-wage one without ever travelling.
Critics describe this as a new form of wage arbitrage. It would carry the conditions of gig work, with piece rates and little security, into jobs that have always had to be done locally. For the worker the robot is not a replacement but a remote-controlled tool, and the employer gains access to the cheapest labour in the world.
For countries such as India, this cuts two ways. It could create a new category of outsourced work, as call centres and data labelling did. It could also mean that the first Indian jobs created by humanoids are low-paid operator roles serving distant customers.
A temporary scaffold?
The industry’s reply is that all this is temporary. Human demonstrations and remote control are scaffolding, it says, to be removed once the robots have learned enough. Every hour of teleoperation is also an hour of training data, so the operators are, in effect, teaching their own replacements.
That may prove true. But the timetable is unknown, and scaffolding has a way of staying up. Self-driving cars have needed remote assistance for far longer than their makers predicted.
In the meantime, a simple rule helps in reading any robot video. Ask whether the task was done autonomously, by remote control, or by a mixture. Honest companies now say. Where a company does not, the safest assumption is that a person was somewhere in the loop.
The popular fear is that robots are taking human jobs. The evidence so far points to something stranger. Before they remove any work, humanoid robots are creating a new kind of it: unseen, modestly paid, and devoted to making itself unnecessary.
