China's humanoid robots aren't smart enough to take your job – yet

Inside a training center in southern China, more than 100 humanoid robots stand in neat rows. A dozen human trainers wearing headsets guide some of them through simple tasks – sorting crates, packaging noodles, making coffee.

The humanoids are slow and awkward. A novice trainer might get them to produce one usable movement in 300 attempts; an experienced trainer, one in 50.

Robot maker UBTech won an $18 million tender from the Guangxi regional government last October to supply the training facility with humanoids and related hardware. The mission: to produce training data for so-called embodied AI – robots able to perceive, decide and act in the physical world. The trainers use headsets to see from the robots' perspective and handheld controllers fitted with sensors to translate their own movements into the robots' actions.

Teaching humanoids to handle labor-intensive work could one day meet an acute need. China's population is aging and shrinking, eroding a critical demographic advantage for its industrial base: a vast pool of workers. And China's military researchers have explored the potential uses of robotics and artificial intelligence in future weapons systems.

The Liuzhou center, which Reuters visited in April, aims to sell robot-training data to factories. But three staffers acknowledged the subsidy-dependent project currently has no clear path to profitability, citing high operating costs and low prices for its data in the nascent market.

UBTech and the Guangxi and Liuzhou governments didn't comment for this story.

The facility captures a central tension in China's humanoid push: Government funding and manufacturing investment are racing ahead of the robots' ability to operate outside of choreographed demonstrations – like the kung fu routines, dances and boxing matches that have made the machines internet sensations.

Though China's robot makers produce world-beating hardware, the humanoids lack the intelligence needed for general-purpose work, limiting their practical uses, a half-dozen industry executives and researchers told Reuters. Critically, they struggle with situations requiring intuition or deviation from programmed routines.

"The robots' IQ is too low," said Tang Wenbin, co-founder and CEO of AI robotics venture Yuanli Lingji, at a March industry panel. "A lot of what we see is dancing disguised as working."

This Reuters examination of China's humanoid boom is based on hundreds of corporate filings and government procurement records; visits to factories, robot-training centers and humanoid showrooms; and interviews with 40 people including executives, investors and researchers. It reveals a widening gap between government ambitions and robot capabilities – resulting in a market where manufacturing capacity outpaces commercial demand, key players rely on subsidies and prices are falling.

The humanoid-robot industry shipped about 20,000 units globally last year, with Chinese manufacturers accounting for 95%, according to BofA Global Research. Gan Xiaobin, an official at China's Ministry of Industry and Information Technology, said in July the country expects to build more than 100,000 humanoids this year. BofA forecasts 1.2 million annual shipments globally by 2030.

The implications of China's supply-chain dominance extend beyond commercial competition. In July, the Trump administration banned imports of new Chinese humanoids, citing the need to protect the U.S. AI industry from threats of disruption and cyberattacks.

Beijing's effort to build intelligent robots follows the industrial-policy blueprint it deployed for electric vehicles and solar panels: Subsidize a strategic sector, build capacity early, and stoke fierce competition to drive down costs. Such investment serves China's ambition to win control of critical industrial technology and displace Western competitors.

But cars and solar were mature industries when China accelerated production. Humanoids, by contrast, remain an early-stage technology with limited domestic or foreign demand and unresolved problems in dexterity, reliability, intelligence and cost, analysts and industry executives say. Government entities are at once major financiers and customers of the same robots – artificially boosting demand.

That all makes China's gamble on humanoids a riskier prospect.

Some analysts question whether a general-purpose machine built in humans' image makes economic sense when robotic arms – purpose-built for specific manufacturing tasks – are cheaper, more reliable and already in wide use.

"The economics do not support widespread deployment today," said Poe Zhao, an analyst and author of the Hello China Tech newsletter. Yet China's capital markets have priced humanoid technology "very aggressively," with investors essentially betting on the robots' future usefulness, he said.

Zhao sees potential in niches such as homes or care facilities, where a human-like machine could use stairs, doors and tools with minimal modification. But in the near term, the market for humanoids "will remain concentrated in dangerous, repetitive and structured work."

China's Ministry of Industry and Information Technology and its State Council Information Office didn't respond to requests for comment about humanoid development.

Kevin Xu, of U.S. investment firm Interconnected Capital, said some founders have characterized China's humanoid sector as a "bubble."

"There will 100% be a consolidation sooner or later," Xu said. "It's just a matter of who and how ugly."

BOOM BUILT ON STATE MONEY

China has more than 150 humanoid-robot companies, according to the National Development and Reform Commission, the top economic planner. That exceeds the number of Chinese electric vehicle brands. The agency, which didn't respond to questions, said in November that companies were crowding into humanoids before the technology and business models had matured.

The country's national, provincial and local governments spent at least $230 million buying humanoids and related items in the first half – up from $62 million a year earlier and $6 million in the corresponding 2024 period, public procurement data show. Reuters reviewed nearly 1,000 fulfilled tenders covering state-entity purchases of robots, training systems, equipment and demonstration projects.

China's industry ministry and its state-assets regulator announced in June that they would require 10 provinces to each identify at least 20 sites for real-world training of humanoids and AI systems.

In April, power utility State Grid announced plans to spend $1 billion on AI-enabled humanoids, dual-arm robots and robot dogs for grid maintenance and substation inspections. Each unit would save $74,000 to $119,000 in annual labor costs and reduce human exposure to hazardous work, state media reported. State Grid didn't respond to questions.

Shenzhen aims to build a $15 billion industry cluster comprising more than 1,200 embodied-intelligence robotics companies by next year. The city didn't respond to a request for comment.

Proponents of the buildout say that it could help Chinese companies create low-cost supply chains and collect robot-training data faster than foreign rivals such as electric vehicle giant Tesla, which aims to go big on humanoids. But the drive exposes investors and local governments to the risk of a market correction.

China's strategy of flooding the market with subsidies, competitors and production capacity is designed to "kick start the Hunger Games," an artificially hypercompetitive market that accelerates learning and innovation, said Lizzi Lee, who studies China's economy at the Asia Society Policy Institute in New York. In other industries, the tactic has depressed prices for years – making it impossible for almost all Chinese and foreign competitors to profit – but ultimately left China with a handful of globally competitive survivors.

"Waste is not an accident in this model" but rather "part of the discovery mechanism," Lee said, noting the inevitably high failure rate. Production overcapacity, however, "is easy to start but hard to unwind."

For now, China's UBTech and Unitree look like potential survivors, many analysts say. Unitree – the world's largest seller of robot dogs and the second-largest humanoid producer – has become a bellwether for capital flowing into the sector. Shares of Unitree soared more than five-fold in their Shanghai trading ​debut on August 19, valuing the company at around $50 billion, before pulling back. Chinese embodied-AI firms eyeing their own IPOs are closely watching Unitree's progress as investors try to distinguish between robot companies with tangible businesses from those selling visions.

Unitree didn't respond to a request for comment about the IPO.

STATE-DRIVEN DEMAND

So far, demand for humanoids has been created by government purchases and subsidies rather than market forces. China's industry ministry and state-assets regulator have pledged preferential policy support for regions and companies with strong robot-training programs.

As with the early EV race, local governments are now vying to create robotics champions. Beijing's E-Town tech zone, also known as Yizhuang, offers robotics firms sales subsidies, free workspace, computing support and vouchers to purchase data, city documents show.

In May, components supplier Lingyi iTech began production at a "super factory" in the district, providing hardware, assembly and testing for robot brands. The facility's first phase set a target of 10,000 embodied-intelligence robots annually, scaling to 500,000 by 2030.

Philip Yang, Lingyi's vice president, during a media tour in May credited municipal support for enabling the plant to go from planning to production in four months. Beijing and Yizhuang officials didn't respond to questions.

Financial disclosures by robotics firms show the risks of policy-linked demand.

Deep Robotics, a Hangzhou-based manufacturer of industrial quadruped and wheeled robots, has disclosed that subsidies accounted for about 42% of its $4.2 million net profit last year. The firm, which sold only four humanoids across 2024 and 2025, didn't respond to questions about the impact of subsidies and the commercial outlook for humanoids.

In the fourth quarter of last year, UBTech won new contracts worth at least $123 million to supply state-owned enterprises with humanoid hardware and related technology, public filings and announcements show. Those deals exceeded the firm's total 2025 humanoid revenue of $122 million. The company, which didn't comment to Reuters, shipped 1,079 full-size humanoids last year.

Competition is driving down prices. Morgan Stanley expects the average price of a humanoid to decline 15% in 2026, before recovering modestly thereafter. The average price of Kuavo, a full-size humanoid made by Leju Robotics, fell 26% year-on-year in 2025 to about $46,000.

Kevin Li, chief marketing officer at Chinese robot-rental platform Sharebot, said dancing robots - which companies lease for promotional events - now rent for $440 to $600 a day, down from roughly $1,500 last year.

PRODUCTIVE FORCES?

China's humanoid race has been built partly on spectacle, with robots boogeying and boxing on television to symbolize President Xi Jinping's "new quality productive forces" – growth built on technological innovation and advanced manufacturing.

The performances mask a fundamental question: whether the machines can generate enough value in the workplace to justify their cost.

Factory managers measure robot performance by uptime and reliability. Conventional robotic arms usually beat humanoids, said Rachel Tan, secretary-general of the Shenzhen Robotics Association, an industry group.

In China's ultracompetitive auto sector, production lines are already dominated by industrial robots that are faster than humanoids and easier to control, two industry participants said.

At Xiaomi's EV plant in Beijing, 91% of the assembly process is automated, the company has said. The body workshop uses over 700 industrial robots for spot welding, riveting and adhesive bonding to build the vehicle frame.

Xiaomi, which declined to comment for this story, is training humanoid robots for tasks such as installing nuts. When Reuters visited recently, the factory floor had few workers.

Geely's Zeekr brand partnered with UBTech last year to test humanoids at its Ningbo factory on tasks including sorting materials, moving boxes and handling components. But for now, Geely told Reuters, robotic arms and automated vehicles "handle all key tasks on the active assembly line."

PICKING PILL BOTTLES FROM SHELVES

Early commercial cases for humanoids are emerging. Humanoid maker Galbot is deploying robots in pharmacies across more than two dozen Chinese cities to address staffing constraints and meet medicine-delivery demand.

At a Beijing pharmacy, Galbot's robot picks products from shelves after receiving digital orders. The robot has a head, arms and torso but glides on a wheeled base. The system uses cameras, including one on the robot's wrist, together with a gripper and suction cup to identify and retrieve items.

Zhao Yuli, Galbot's chief strategy officer, said a humanoid can pick and prepare a product in about one minute. A typical pharmacy stocks 5,000 to 6,000 products in a 30-to-50-square-meter area, Zhao said, creating a repetitive selection task that can be automated. The robots' success rate exceeds 95%, Galbot says, with most failures involving difficulty grasping an item rather than selecting the wrong product.

The company said it hadn't timed the robots against human workers, but believed the machines' ability to work continuously could make them equally or more efficient overall — even if humans are faster at individual tasks.

The AI brains of humanoids differ from those in chatbots. Chatbots such as DeepSeek and Kimi are powered by large language models that process huge amounts of text. A factory humanoid needs to process vast quantities of visual data and convert what it sees into physical action: Gripping a piece of fabric; calibrating force on a screw.

Companies call these vision-language-action models. They are much harder to train than LLMs, and their development is at an earlier stage, because they require thousands of hours of exposure to real-world scenarios. The problem, researchers and executives say, is these models struggle to adapt: A robot may execute a task correctly after training on one workstation but fail when the object, lighting, angle, surface, timing or equipment changes.

"The problem goes from the last 10 centimeters to the last 1 centimeter, even 1 millimeter," said Jia Baoxiong, a research scientist at the Beijing Institute for General Artificial Intelligence.

DATA BOTTLENECK

China's humanoid campaign relies on a simple equation: More robots in training equals more data.

A phase of overproduction is necessary to improve embodied AI models, said Robert Wu, a Shanghai-based tech investor and CEO of Baiguan, a market data and research firm. "To build those models, you first need data; and to get data, you need a massive installed hardware base out there collecting it."

The sector has roughly 500,000 hours of high-quality training data, but needs 100 million hours for robots to achieve true physical intelligence, said analyst Poe Zhao, citing industry estimates. "That is an orders-of-magnitude deficit that neither capital nor volume can close quickly," he said.

Other efforts focus on improving the humanoid brain. Chinese startup X Square Robot said at a company event in April that it has been training its models on data collected in 100 households, exposing systems to varying lighting, clutter and unfamiliar objects that can disrupt robots trained in controlled conditions.

SURVIVING THE POP

Many robotics founders and investors predict consolidation could begin in late 2026 or 2027, as they expect Chinese officials to rein in subsidies and focus support on the most credible players.

"Every single one of these companies knows that the ultimate commercial goal is still a long way off," said Wu, the investor. "That is why they are desperately trying to grab as much capital as possible right now – to buy themselves enough time to survive the bubble's pop."

Already, Chinese component prices are falling so quickly that competitors abroad struggle to keep pace, according to BofA Global Research's Ming Lee.

Tesla illustrates the gap. Its Optimus humanoid isn't yet sold publicly. CEO Elon Musk's stated target price of roughly $20,000 is a long-term goal at large-scale production. Morgan Stanley, however, last year estimated the low-volume hardware bill of materials for an Optimus Gen 2 model at $50,000 to $60,000.

Tesla didn't respond to a request for comment.

China's robotics industry will likely maintain its overall global advantage through a potentially painful consolidation, some analysts say. Even if bankruptcies sweep through humanoid manufacturers, China's supply chains will persist, potentially locking foreign competitors into dependence on Chinese hardware.

That dynamic was on display in South Korea in July at RoboCup 2026, where teams consisting of humanoid robots played soccer against each other. Previously, teams brought their own robots to the annual international robotics and AI competition, said Kim Ho-joung, a manager at South Korean humanoid manufacturer AeiROBOT. Now, almost all use Chinese robots, he said.

Despite China's infrastructure advantage, humanoid robots remain too slow and error-prone for broad commercial adoption.

At the Liuzhou training center, Reuters observed a humanoid struggle to carry crates. The robots can complete simple tasks at just 20% of a human's speed or output, one staffer said.

But they might catch up in two to three years, the staffer added.