The Hidden Human Labor Behind Training Humanoid Robots

The push for humanoid robotics has brought with it an increase in human labor dedicated to training, supervising, and even operating these machines—yet much of this work occurs out of the public eye. Despite claims of rapid advances in ‘physical AI,’ a term referring to artificial intelligence that powers robots with human-like physical abilities, these systems continue to rely heavily on input from human workers.

Jensen Huang, CEO of Nvidia, recently forecasted a new era of physical AI in which machines would perform everyday tasks in homes, workplaces, and public spaces. Demonstrations of robots assembling cars or performing household chores suggest significant progress, but the underlying human effort tells a more complex story.

A critical aspect is the creation of training data. Using techniques such as imitation learning, robots are shown how to complete actions—sometimes by humans performing those tasks repeatedly, while sensors record every movement. In Shanghai, one worker reportedly wore a virtual reality headset and an exoskeleton to repetitively open and close a microwave, gathering data for a robot to emulate. Similar efforts are underway in North America, where robotics company Figure, in partnership with Brookfield, plans to record large-scale data from real homes to improve machine performance. (Figure declined to comment on specific details.)

Industry insiders anticipate that, much as people’s online content was used to train large language models, everyday human movements may become fuel for the next wave of robotic learning. Aaron Prather, a robotics specialist, described recent projects where logistics workers wore motion sensors while moving boxes to supply data for training autonomous systems. Such practices are expected to become widespread, turning manual laborers into large-scale data contributors.

Tele-operation—where humans remotely control robots over the internet—remains another less-visible form of labor supporting these systems. For example, 1X, a startup behind the Neo humanoid robot, revealed that customers may rely on company tele-operators to remotely assist with challenging household tasks, should the robots encounter problems. While the company seeks customer consent for remote operation, the practice raises privacy issues and suggests that full autonomy is still out of reach. Moreover, using remote labor to operate robots in homes presents new forms of wage arbitrage, drawing parallels with existing gig work models.

Analogous patterns have emerged in content moderation and AI data labeling, roles often outsourced to low-wage workers, sometimes involving distressing material. Despite optimism that AI models will soon be self-teaching, substantial human supervision, correction, and intervention remain essential.

The invisibility of these human support structures does not mean AI is ineffective or misleading, but it does contribute to public misunderstanding. Overestimating robot capabilities may encourage unsafe behaviors or excessive trust—for example, as seen in recent legal challenges over autonomous vehicle claims.

Greater transparency from robotics companies about the extent of human involvement is likely necessary. Otherwise, society risks conflating hidden human labor with genuine machine intelligence, potentially misjudging the readiness and reliability of emerging technologies.

technologyreview.com