The appearance of DeepSeek-R1, a Chinese AI model that seems to rival OpenAI’s latest offerings far more cheaply, shocked markets this week and erased $1 trillion from U.S. stock values. This event underscored the stakes of America’s technology race with China—and how close that race is. But as well as competing over AI models such as ChatGPT and DeepSeek-R1, these two tech superpowers are now also racing for a new prize: robots shaped like humans, with a head, torso, arms, and (often) legs. Such humanoid robots are central to the future plans recently announced by Jensen Huang, CEO of chipmaker Nvidia. Huang’s vision has led Nvidia’s rise to become one of the world’s biggest listed companies. Elon Musk correctly anticipated vast markets for space launch and electric vehicles—and Musk now predicts that the long-term value of Tesla’s humanoid robots “will exceed that of everything else at Tesla combined” and make it a $25 trillion company. Meanwhile, Chinese industrial policy is pouring a firehose of start-ups into humanoid robots.
Advances in generative artificial intelligence since 2022 have turbocharged the development of humanoid robots, and this is accelerating. Twenty-seven humanoid robot models reportedly debuted at Beijing’s World Robot Conference in August 2024. A few months earlier, Huang had announced a new foundation model—the underlying model on which specific uses are built—developed for controlling humanoid robots. Humanoid robots aren’t yet viable in many real-world environments but have begun operating in Amazon warehouses and factories for Mercedes-Benz and BMW. Goldman Sachs Research anticipates a market worth $38 billion by 2035, and Citibank estimates 648 million humanoid robots by 2050.
The appearance of DeepSeek-R1, a Chinese AI model that seems to rival OpenAI’s latest offerings far more cheaply, shocked markets this week and erased $1 trillion from U.S. stock values. This event underscored the stakes of America’s technology race with China—and how close that race is. But as well as competing over AI models such as ChatGPT and DeepSeek-R1, these two tech superpowers are now also racing for a new prize: robots shaped like humans, with a head, torso, arms, and (often) legs. Such humanoid robots are central to the future plans recently announced by Jensen Huang, CEO of chipmaker Nvidia. Huang’s vision has led Nvidia’s rise to become one of the world’s biggest listed companies. Elon Musk correctly anticipated vast markets for space launch and electric vehicles—and Musk now predicts that the long-term value of Tesla’s humanoid robots “will exceed that of everything else at Tesla combined” and make it a $25 trillion company. Meanwhile, Chinese industrial policy is pouring a firehose of start-ups into humanoid robots.
Advances in generative artificial intelligence since 2022 have turbocharged the development of humanoid robots, and this is accelerating. Twenty-seven humanoid robot models reportedly debuted at Beijing’s World Robot Conference in August 2024. A few months earlier, Huang had announced a new foundation model—the underlying model on which specific uses are built—developed for controlling humanoid robots. Humanoid robots aren’t yet viable in many real-world environments but have begun operating in Amazon warehouses and factories for Mercedes-Benz and BMW. Goldman Sachs Research anticipates a market worth $38 billion by 2035, and Citibank estimates 648 million humanoid robots by 2050.
Yet a robot is any machine that can perform a complicated series of tasks automatically, and robots can take many forms—such as robot arms in factories, self-driving cars, or military drones—so what are the advantages of a humanoid? Humanoid robots are a glittering prize for two reasons, which together promise a mutually reinforcing spiral. One is the huge potential market for robots that can use our human tools and function in our human environments. Second, as artificial intelligence butts up against limits to available data, humanoid robots offer a route to transform AI itself. With this huge prize on the line, it’s no wonder the world’s two superpowers, equally matched competitors, are seeking out every edge in the contest for human forms.
Giving robots humanoid form—or at least aspects of it—opens up vastly more possibilities for those robots to act usefully in the human world, in which we have spent millennia, as well as trillions upon trillions of dollars, making tools and environments for humans shaped like us. A world full of stairs, tables, screwdrivers, medical instruments, and so on. A robot would be very useful if it moved around my Victorian house in London and used my tools to clean or cook. The same with workplaces such as factories, hospitals, elder care homes, or battlefields. Moreover, robots shaped to operate throughout our world can better accompany us as we go about our tasks, to collaborate with humans in teams.
The human form also inspires new ways to make useful robots. Walking like a human with straight legs, for example, saves energy compared with typical robots that walk with bent legs. Human hands are awesome. Evolution gave humans fantastic capabilities, which is why so much of AI is benchmarked against human capabilities, from the Turing test of language to computer vision or the boardgame Go. Earlier generations of robots drew inspiration from simpler creatures such as insects or dogs—and today’s AI advances in areas such as language and planning make the human form a ripe target.
Mass deployment of humanoid robots won’t be immediate. Advances in AI perception took years to reach mass scale in Amazon Alexas. Robotaxis spent years developing in U.S. and Chinese cities, and only in 2024 did Google’s Waymo reach an inflection point to surge from 1 million to 4 million passenger trips. But the humanoid form is so useful that there are huge addressable markets for them as the technology develops. And that is only half of the spiral that humanoid robots promise—the second is to fuel a new leap in AI itself.
AI spent decades in the doldrums until a huge leap in computer vision in 2012, and that required a big new dataset of visual images to train the AI. 2022 saw a huge leap in generative AI with ChatGPT, and training that AI again required a huge leap in data, which in that case involved much of the internet. Both advances also needed enhanced computer power and computational techniques, but the data was crucial. Today, we have exhausted all the world’s easily accessible data for training models, so where can the next big leap in data come from? Simulating data helps, but we also need data grounded in reality. Humanoid robots can help provide vast new data, linked to reality, for learning how to act in the physical world.
Robots are physically embodied. Today’s AI can give robots vision-language-action models that can take in text (like in ChatGPT), plus data from the robot’s physical environment (e.g., via cameras or microphones) and from internal sensors (e.g., of joint positions in a robot’s hand). It’s incredible how much data even a single human infant gets from their “external sensors”—a recent study used video and audio from a head-mounted camera on a single infant, and that data alone enabled an AI to learn many words and concepts. Adding the effects of an infant or a robot’s actions gives even more useful data about how the world works. And combining all these types of data can reduce hallucinations because when you stub your toe, for example, that’s a collision with reality.
Humanoid robots gain additional advantages. They can learn from the actions of Earth’s most remarkable intelligences: humans. That’s why Nvidia’s Project GR00T aims to develop AI models that help humanoid robots learn by observing human demonstrations and by having human teleoperators help robots practice actions. How humans perform tasks can give humanoid robots the “right answers” from which to train. Such help is vital because of Moravec’s paradox: Tasks thought difficult for humans are often easy for AI (e.g., chess), but tasks thought easy for humans (e.g., putting a shirt on a hanger) can be very hard for AI. The right answers for actions such as putting shirts on hangers seem easy to humans, but robots can learn a lot from how humans succeed at such actions. If robots share our human form, it will also be easier for us to teach the robot: to explain how we perform actions with our bodies and to provide robots with feedback on their efforts that can help them learn.
Such learning can happen at mass scale with millions of humanoid robots, bringing together both halves of the mutually reinforcing spiral that makes humanoid robots such a glittering prize in the global tech competition. More robots interacting with humans leads to more data from which their AIs can learn, which leads to better AI that in turn enhances the robots so they take on more jobs, which leads to more data from which their AIs can learn, and so it spirals onward.
In the global race to win the prize of dominance in humanoid robots, China and the United States have different strengths—and each relates to a different half of this spiral involving mass manufacturing and AI learning.
China’s big edge is manufacturing at scale. China is the world’s sole manufacturing superpower, with production exceeding the nine next largest manufacturers combined. China dominates some key robot markets, such as drones, where in 2023 DJI alone supplied 70 percent of the world’s drone users. China seems equal to the United States in robotaxis. And China’s huge EV companies are investing heavily in driver-assistance software to make their cars, as one Chinese auto executive described, “a robot on wheels.” Robotics in general was identified as one of Chinese President Xi Jinping’s “new quality productive forces,” and in 2022 China installed more than half of the world’s industrial robots.
Humanoid robots were identified by China as a key area for technological competition in 2023, when the Ministry of Industry and Information Technology released its “Guiding Opinions on the Innovation and Development of Humanoid Robots.” This year China seeks to establish a world-class innovative ecosystem for humanoid robots, and by 2027 it wants to integrate humanoid robots into manufacturing supply chains, use them at scale, and expand humanoid robot use throughout society.
The United States’ big edge is in the most cutting-edge technologies at companies such as Nvidia, in its hub of start-ups, and at universities—and although America lacks China’s manufacturing scale, this could help it build better AI robot learning. A humanoid’s software “brain” accounts for roughly 80 percent of its value, and Nvidia chips still seem preferred for much cutting-edge AI in China. A U.S. company built ChatGPT, and U.S. researchers still push China into second place for publishing top-cited AI research. Allies are also key, with Britain for example publishing the third-most top-cited AI research, while Switzerland, Germany, and Japan manufacture many of the world’s industrial robots.
But although China and the United States are evenly matched today, that can change. America pioneered much in the industries that manufacture semiconductors and industrial robots, for example, yet eventually fell far behind. China’s new DeepSeek-R1 suggests America’s edge in AI is hardly unassailable. So, what can Washington do?
Competition to lead in humanoid robots may be the most consequential technological race of the next decade, but except for the rare few such as Nvidia’s Huang, most people don’t even know this race exists. A first step is greater awareness of this vital race among U.S. policy communities. Next, a better understanding of the challenge will help policymakers navigate the trade-offs needed to win this race, as inevitable political pressures around job losses, privacy, and political control affect these developing industries. Policy levers such as industrial strategy or tariffs (depending on political preferences) can also help protect the development of these vital new industries through their difficult early stages, in order to match Chinese efforts.
The United States must not lose its advantage in the most cutting-edge research, but much of what is needed to compete in this race also requires a return to excellence in manufacturing at scale. And this itself reveals an intriguing change in our relationship with AI. Increasingly, AI will leave the virtual realm of TikTok, X, or Instagram feeds, and enter the hard reality of the physical world around us.
The post The Global Race for Humanoid Robots Heats Up appeared first on Foreign Policy.
