Thirteen years ago, Gill Pratt ran the DARPA Robotics Challenge — the Pentagon's attempt to build a robot that could drive, climb rubble, and turn a valve when a Fukushima-style disaster knocked out its operators. The robots failed spectacularly, often comically. Pratt spent the next decade watching humanoid robotics cycle through waves of hype and disappointment.
Now, as CEO of the Toyota Research Institute, he says the moment is finally different. What changed, Pratt told IEEE Spectrum, is not the body — researchers have been building humanoid-shaped robots for decades. What changed is the brain. Diffusion policy, the AI technique TRI pioneered two years ago, and the larger behavior models built on top of it have cracked the data problem that stalled the field for a generation. Every robotics demo you've seen in the past year runs on some form of diffusion-based training that traces back to his lab.
But here is what Pratt's bullish interview leaves out: the entire humanoid robotics revolution runs on $15-an-hour gig workers in Nigeria, India, and Argentina recording themselves folding laundry.
Scale AI, the data infrastructure company, announced it had gathered more than 100,000 hours of human movement data for the sector. Companies including Micro1 are paying workers in lower-income countries to perform household and industrial tasks while sensors capture every motion. One worker in Nigeria folds shirts, stacks shelves, and opens drawers so that a machine can learn what those actions look like. Ken Goldberg, a robotics professor at UC Berkeley, put the data problem in context: large language models were trained on text that would take a human 100,000 years to read, and humanoid robots may need even more data, because controlling robotic joints is more complex than generating text. Robotics companies are now spending more than $100 million each year buying real-world training data from companies like Micro1 and Scale AI, according to Micro1's CEO.
The dependency ratio Pratt cites as the societal driver — more retirees, fewer young workers — is real and worsening. Elder care is a genuine market. But the hype cycle is running ahead of the reality. Pratt himself acknowledges that humanoid robotics is approaching the peak of inflated expectations and predicts a trough of disillusionment is coming. Nobody, he said, is thinking deeply enough about the gap between what he calls system one and system two cognition: the fast, instinctive pattern-matching today's robots do, versus the slow, deliberate reasoning that would let them handle genuinely novel situations.
Rodney Brooks, the iRobot cofounder whose Roomba vacuumed millions of homes before today's humanoid companies shipped anything, has been blunter. His advice: stay at least three meters from a full-size walking robot. The hand dexterity problems and the walking stability problems that plagued the DRC in 2012 have not been solved. They have been papered over with better demos.
The factory question illustrates the gap. Humanoid companies are racing to deploy their robots in manufacturing environments. Pratt noted something unexpected in his IEEE Spectrum interview: factories are flat. Wheels work fine on flat surfaces. It is very weird, he said, to see so much focus on legged robots in factories when wheels would do the job.
The money is real. Humanoid robotics startups raised nearly $14 billion in 2025, up from $8.2 billion in 2024, even topping the $13.1 billion raised in the peak venture funding year of 2021. Apptronik closed a $520 million extension at a $5 billion-plus valuation, bringing its total raised close to $1 billion.
The research progress is real. The ghost workers in Nigeria and India doing $15-an-hour data work are also real, and they are part of what makes the progress possible. Whether that progress translates into robots that work reliably next to humans, rather than alongside them in carefully choreographed videos, is the question the next several years will answer. Pratt has seen this movie before. He says he is hoping for the best while preparing for a bumpy road.
