Gasgoo Munich-"The latest data shows that China accounts for 97% of global humanoid robot exports and manufacturing."
Jin Bing, a former deputy director and first-class inspector at the State Post Bureau’s policy and regulations department, cited those figures on September 9 during the Converged Intelligent Industry Development Conference (2026).
For years, the industry was still proving whether humanoid robots could even be built. Walking, running, jumping, and the ability of dexterous hands to grasp objects were often the most direct calling cards for a robotics company.
By 2026, however, the questions keeping the industry awake are shifting.
Data presented by Jin shows global shipments of general-purpose humanoid robots reached about 18,000 units in 2025, while China is expected to ship 62,500 units domestically in 2026. Production lines capable of 10,000 units are already up and running, and deliveries by leading players have quickly surged from hundreds to thousands—and beyond.
Image source: Screenshot from Deputy Director Jin Bing’s presentation
Humanoid robots are transitioning from laboratory curiosities into genuine industrial products.
For Chinese companies, figuring out how to manufacture these robots remains difficult—but it’s no longer the only problem. A more pressing question looms: Once built, how many of these robots will actually stay on the factory floor, in the warehouse, or in other real-world environments for the long haul?
The mass production hurdle is being cleared
"The year of mass production" has been a buzzword in the humanoid robot sector for the past two years.
Jin identifies 2025 and 2026 as the dawn of mass production, with 2027 through 2028 marking a gradual shift toward large-scale adoption.
That shift is already visible in the data. The State Taxation Administration previously disclosed that revenue for China’s embodied intelligence sector climbed 22.4% year-on-year in the first five months of 2026. Robot manufacturers saw a 30.1% jump, while AI algorithm and software integrators rose 24.5%, and system integration firms grew 27.9%.
Beyond the OEMs, the upstream supply chain is heating up as well.
Motors, batteries, reducers, lead screws, sensors, and structural components—along with the vast supplier network that once served the automotive, consumer electronics, and industrial automation sectors—are all pivoting toward humanoid robotics. The manufacturing prowess China has accumulated over decades is rapidly absorbing this new industry.
Jin summarizes China’s current edge as “mass production, a complete supply chain, and scenario validation.”
The dexterous hand is a prime example. Some domestic five-fingered hands have already caught up with overseas products in metrics like degrees of freedom, with mass-production costs running at roughly a third of those abroad.
Deputy Director Jin Bing’s presentation
The playbook is familiar. Scale up first, drive down costs through supply chain and manufacturing efficiency, then push for even greater volume. It’s the same trajectory the consumer electronics and new energy vehicle industries followed.
But humanoid robots have one key difference.
Building more doesn’t automatically create more demand.
A factory capable of producing 10,000 robots a year doesn’t mean there are 10,000 mature job openings waiting for them. The real challenge for humanoid robots is shifting from the assembly line to the floor beyond it.
Robots are shifting from ‘performance’ to ‘work’
In June, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special initiative for 2026 focused on real-world training for humanoid robots and embodied intelligence.
One notable goal is moving robots from tech showcases into actual production and living environments—switching on “work mode.” The mandate sets a target: by the end of 2026, establish over 100 high-value application scenarios and drive the capability to deploy units at a scale of 10,000.
"Performance" and "work" may differ by only two characters, but behind them lies a completely different set of evaluation criteria.
How fast a robot runs, whether it can backflip, or if it can dance certainly demonstrates hardware and motion control capabilities. But inside a factory, buyers have different concerns.
How long can it run continuously? How high is the success rate after repeating a motion 10,000 times? If a material shifts by a few centimeters, can it still grasp it? How long does it take to retrain for a new workstation? Who handles repairs when it breaks down? And most importantly: when you run the numbers, is it actually more cost-effective than human labor or traditional automation?
Right now, those questions don’t have complete answers.
In terms of shipment structure, scenarios like entertainment, commercial performances, and educational data collection still account for a significant share. Smart manufacturing and warehousing logistics are growing, but they haven’t yet become the dominant mainstream.
So the current “year of mass production” is really just a beginning. In the past, there weren’t enough robots in real environments, so many problems stayed theoretical in the lab. Now that numbers are rising, the industry finally has a chance to prove at scale whether these machines can actually do the job.
Wang Xingxing, founder of Unitree Robotics, has raised a similar point. A robot can achieve high success rates after intensive training in a fixed scenario, but its performance can quickly degrade the moment the environment or the object changes.
The real challenge isn’t whether it can do this one task—it’s whether it can still do the task in an environment it’s never seen before.
That’s a key difference between humanoid robots and traditional automation. Traditional robotic arms can be endlessly optimized for a specific process. Humanoid robots, however, are counted on to handle tasks that aren’t standardized—and that change constantly.
The ideal is clear; reality needs time.
Scenarios are plentiful, but they must first become ‘valid data’
Solving the generalization problem ultimately circles back to data.
Yao Maoqing of Mifeng Technology recently noted a fundamental difference between language models and embodied intelligence regarding data sources, in an interview with Gasgoo Embodied Intelligence.
Much of the knowledge required for language models already exists on the internet. But robots need to perform physical tasks, and many of those real-world skills and experiences simply aren’t online.
"The volume is four or five orders of magnitude smaller than language models," Yao said. "Second, authenticity and diversity aren’t just buzzwords. You have to go into stores, factories, and hotels, and collect data task by task using wearable devices. Only then can you cultivate an embodied generalist—someone who can infer and apply knowledge only after having seen it."
That’s also why China’s scenario advantage is often cited.
China has a vast number of factories, warehouses, malls, hotels, and logistics hubs. For robotics companies, these are potential markets—but they hold another value: they are places where robots can constantly gain real-world experience.
If this cycle can be established, more robots entering diverse real-world scenarios will generate more data. That data trains models, boosting task success rates and generalization, which in turn helps robots enter even more scenarios.
Jin summarizes it as “trading scenarios for technology, reducing costs through mass production, and filling gaps with a comprehensive system.”
But having “many scenarios” doesn’t mean the data is automatically valuable.
Yao is blunt: “You can’t judge a data company’s quality just by volume. Blindly collecting 1 million hours is honestly easy—just repeat the same task over and over. But it has almost no value.”
In his view, data quality depends on how many industries, scenarios, and objects are covered; whether the data is diverse enough; if a complete quality validation report can be provided; and ultimately, whether the client actually signs off on it.
This is a very real problem for the data industry right now.
Claims of “millions of hours” are becoming common, but the number itself doesn’t mean much. One million hours spent repeating the same motion is not the same as one million hours covering different environments, objects, operational methods, and edge cases.
Of course, scale still matters.
Yao notes that tens of thousands of hours of real-world robot data is now a standard requirement for model training, but collecting that much isn’t easy—it might take thousands of robots running continuously for six months.
That’s precisely why different approaches—like teleoperation, body-less data collection, and simulation—are all emerging simultaneously.
The coming industry race won’t be about who first announces “100 million hours of data.” It will be about who can acquire truly useful data more cheaply and reliably—and turn it into capability.
If this step isn’t achieved, China’s rich scenarios will remain just that—abundant resources.
The more robots are built, the clearer the weaknesses become
As production volumes ramp up, the industry’s problems are becoming increasingly specific.
Jin highlighted several areas that still need breakthroughs, including high-end RV reducers, ultra-sensitive tactile electronic skin, high-end force sensors, dedicated computing chips for embodied intelligence, underlying simulation engines, and algorithms for physical interaction in general scenarios, as well as long-term environmental adaptability.
Some are traditional “choke point” components, while others are new problems exposed only after embodied intelligence enters the real world.
Take tactile and force sensing, for example.
For a robot that only needs to walk on a stage, these capabilities might not be urgent. But once it has to screw in bolts, grasp soft objects, or handle precision parts, the exact force applied at the end effector—and whether the object is slipping—directly determines success or failure.
The same goes for simulation.
Robots can’t test every skill on real hardware—it’s too costly and inefficient. But if there’s too large a gap between the simulation and the real world, the trained capabilities won’t transfer easily to actual robots.
Standards issues are also surfacing.
Since the beginning of the year, standards for training grounds, data collection, general requirements for complete robots, and interfaces for dexterous hands and actuators have been advancing. The industry isn’t just solving “do we have robots?” anymore—it’s building the infrastructure for “how do we scale them?”
This is critical.
A robotics company can throw dozens of engineers at adapting a project for a single factory. But if moving to the next factory means re-collecting data, retraining models, and rebuilding interfaces, then “scaling up” just ends up being a pile of custom projects.
Image source: Screenshot from Deputy Director Jin Bing’s presentation
What humanoid robots really need isn’t “a production line that can build 10,000 units.” It’s the ability to take the same product and the same capabilities and rapidly replicate them across dozens, hundreds, or even more customers.
Only then does it truly become an industry, rather than just a project.
From hundreds to thousands, and then to 10,000-unit production lines, China’s manufacturing scale for humanoid robots will keep climbing.
But the faster manufacturing runs, the harder it becomes to avoid another question: can these robots actually stay on the job for the long term?
A cheap robot that only does demonstrations is still expensive. A robot that isn’t cheap—but can work stably for over ten hours a day, replacing repetitive, dangerous, or hard-to-fill roles and paying for itself in a few years—that robot has real commercial value.
So, the “year of mass production” in 2025–2026 may simply be the year humanoid robots truly begin to face the market’s test.
China is getting better and better at building robots.
But the next phase of competition won’t happen on the assembly line.
