China's Humanoid Robot Bet: Why One Researcher Says Doubters Will Be Left Behind
When Unitree, China’s best-known maker of humanoid and quadruped robots, saw roughly 200 billion yuan (about $28 billion) wiped off its market value, critics were quick to call the country’s embodied-AI boom a bubble. But in a lengthy podcast interview, Dr. Allen Yang, a robotics researcher at UC Berkeley, pushed back hard: the people betting against embodied intelligence, he suggested, may be the ones without a future.
Embodied AI is the term for artificial intelligence that operates through a physical body — a robot that walks, grasps, balances and reacts to the messy real world rather than merely generating text on a screen. China has poured money into the field, with startups, listed manufacturers and university labs all racing to build machines that can work in factories, warehouses and eventually homes.
Yang’s core argument is about maturity, not hype. Today’s robots, he says, are excellent at following a path that humans have already defined. The next leap comes when a machine can be given a goal and figure out the route itself. That distinction — between automation and genuine agency — is what separates a useful tool from a transformative one.
The bottleneck is not hardware. It is data. In another interview, Yao Maoqing, a veteran of the embodied-AI industry, described a gap of tens of millions of hours: robots need to learn from far more physical experience than anyone has yet collected. Companies are steadily raising their data budgets, competing to gather footage and sensor readings that are richer and more varied, because a robot that has only ever practiced in one warehouse will fail in the next one.
The economics are beginning to shift. An earlier wave of AI spending went mostly into chips — the GPUs that train large models. As AI agents move from answering questions to taking actions, the money is spreading into servers, sensors and the physical infrastructure that lets software touch the world. In other words, a popular AI application now pulls along a much wider chain of hardware businesses.
That chain is precisely where China’s advantage lies. The country already dominates the manufacture of electric motors, batteries, sensors and precision components, the same building blocks that humanoid robots require. A robot is, in many ways, an electric vehicle that walks — and China has spent a decade learning how to build those at scale, cheaply and quickly.
None of this guarantees success. Researchers openly acknowledge that the excitement around self-improving AI systems contains a speculative element, and that many well-funded robot companies will not survive. The value question is also unresolved: executives admit that if AI investment is not tied to a deep understanding of real business needs, it can become waste.
Still, the direction of travel is clear. Unitree’s share-price swings reflect investor nerves, not a verdict on the technology. The more telling signal is that the money has not stopped flowing — into data collection, into components, into the unglamorous groundwork that has to be laid before a machine can walk into a factory and simply get to work. The race, as Yang frames it, is not about who can build the flashiest robot. It is about who can teach it to find its own way.