Over the past year, announcements surrounding humanoid robots have been coming almost nonstop. Procurement contracts worth tens of millions—or even hundreds of millions—of yuan have appeared one after another.
According to the2026 Humanoid Robot Industry Development Report, more than 40,000 humanoid robots were shipped in China during the first half of the year, accounting for approximately 97% of global shipments.
Yet during the World Robot Conference,People’s Dailypoured cold water on the hype: genuinely commercial, large-scale production orders for humanoid robots remain relatively scarce.
At first glance, the two statements seem contradictory.
If orders are everywhere, why is genuine market demand still considered limited?
The answer lies inwhat these orders actually contain.
1. Beneath the Boom in Orders Lies a Hidden Data Story
Consider one pattern that has repeatedly appeared across the industry.
A company forms a joint venture or subsidiary with other entities, and that subsidiary then purchases robots from the company itself. Some of these transactions effectively amount to money moving from one pocket to another. Others are merely framework agreements, with no firm delivery schedule attached.
But perhaps even more telling is the rise ofrobot data-collection centers.
In one major segment of the market, a data-collection ecosystem reportedly worth RMB 1.4 billion has emerged, with nearly 60% of the funding flowing into data-collection facilities.
The business model is relatively straightforward: a robotics company sells robots to a data-collection center and receives payment. It then turns around and purchases training data from the same center.
The money makes a full circuit.
The robot itself may barely move.
From the outside, this can look like a capital game. But for people inside the industry, there is another interpretation.
These seemingly circular orders are, in essence, a way of spending money to acquire one increasingly valuable asset:
data.
2.The Real Bottleneck Isn’t the Robot. It’s the “Textbook.”
Why have humanoid robots struggled to enter factories and homes at scale?
It is not simply because engineers cannot build capable joints or because manufacturing costs remain too high.
The bigger constraint isdata.
Industry estimates suggest that training a robotic “brain” toward human-level capabilities could require billions of hours of real-world operational data.
Yet as of early 2026, the amount of high-quality physical interaction data available globally was estimated at only around 500,000 hours. Some industry observers have argued that the available volume does not even reach one-tenth of the tens-of-millions-of-hours threshold needed for meaningful progress.
So why not simply rely on simulation?
The problem is that simulation environments are inherently idealized, while the physical world is full of variables.
A slight change in lighting. A cup moved half a centimeter. An object that is 50 grams heavier.
Any of these seemingly insignificant differences can cause a model to fail.
The gap between simulation and the real world cannot simply be closed by making models larger. It requires real-world robot data that covers enough of these variables and edge cases.
That is whererobot data-collection facilitiescome in.
Inside controlled environments, operators recreate scenarios such as factories, pharmacies, warehouses and retail spaces. Real robots repeatedly perform tasks such as grasping, moving, sorting and assembling objects.
Every physical interaction is recorded as data.
That data is then fed back into the model, gradually improving its accuracy and robustness.
3. Data-Collection Facilities Have Eased the Data Shortage—But They Have Also Created a “Well Each Company Drinks From Alone”
Data-collection facilities have undoubtedly helped address the shortage of real-world robot data.
But they have also created a new bottleneck of their own.
The most fundamental problem is the lack ofdata standards.
Research by AIoT Star Map Research Institute found that companies currently use different robot platforms, sensor calibration methods and definitions of actions when collecting data. As a result, the datasets they produce often use incompatible formats and cannot be easily transferred across robot platforms or manufacturers.
Centralized data collection also faces three major challenges.
First, the scenarios are often too artificial.
In a closed training facility, lighting may remain constant and floors may be perfectly level.
The resulting dataset can look more like data from a film set than data from the real world.
Second, the economics are difficult to sustain.
A professional teleoperator may work an eight-hour shift, but only two or three hours of that time may ultimately produce genuinely useful ground-truth data.
Third, the data remains tightly tied to a specific robot platform.
This makes reuse across different scenarios and robot embodiments extremely limited.
The result is a fragmented ecosystem.
A company may spend heavily to collect its own dataset, only to find that the data is useful primarily for its own robot. Move the same dataset to another company’s hardware, and much of its value may disappear.
Hundreds of training facilities may therefore create the appearance of a booming industry.
But underneath the surface, companies are effectivelydigging their own wells and drinking their own water.
The investment is duplicated, while the industry gains little ability to build on one another’s work.
Even the data-collection business itself is already undergoing a shake-up.
According to industry sources, one data-collection facility ceased operations less than 18 months after launch, with the original operator removing its own equipment. The reported reason was that the teleoperation approach had taken the wrong direction, resulting in data quality and accuracy falling short of expectations.
4. The Key to Mass Deployment Lies in Two Things That Have Yet to Be Fully Solved
For humanoid robots to move from exhibition floors into large-scale commercial deployment, companies cannot simply continue collecting data independently.
Two major variables need to fall into place.
1. Cross-Platform Data Standards
The first is the development of standards that allow data to move across different manufacturers and robot embodiments.
In July 2026, the Ministry of Industry and Information Technology’s TC8 established WG7, a working group focused on humanoid robot data standards.
At a working-group meeting, representatives from the lead organization noted that while the industry faces a shortage of real-world interaction data, the lack of standards may be an even more serious problem than the shortage of data itself.
Earlier, in April 2026, the world’s first international standard for humanoid robot datasets was formally approved as a new project under ISO/TC 299, the international committee for robotics, with China taking the lead and several Chinese companies, including Digital China, participating extensively in its development.
Meanwhile, China’s Ministry of Industry and Information Technology has also been soliciting public feedback on theGuidelines for Building the National Humanoid Robot Industry Standard System (2026 Edition), which specifically addresses the full data lifecycle—from collection and processing to storage and circulation.
Once these standards are established, robot data could begin to function more like electricity does today: something that can flow across systems rather than remaining locked inside a single company.
Data collected from an assembly line using Robot A could potentially be transferred to Robot B and put to use with far less adaptation.
That is when data becomes a genuine industry-wideforce multiplier.
2. The Maturation of World Models
The second variable is the development ofworld models.
Traditional large language models predict what comes next in language.
World models aim to predict something much more difficult:
what action should come next—and what physical consequences that action will produce.
There is an important feedback loop here.
The evolution of world models also depends heavily on data.
As massive volumes of high-quality, cross-platform real-world data become available, world models can become significantly more capable of understanding and predicting physical environments.
As the models become stronger, robots become better at generalizing across real-world scenarios.
As robots become more capable, commercial deployment accelerates.
That creates a positive feedback loop:
More data → better world models → stronger robot generalization → broader deployment → more real-world data.
Conclusion
So, back to the question we started with:
Are the orders dismissed as “money moving from one pocket to another” simply a bubble?
In the short term, there is certainly some speculation and inflated activity.
Related-party transactions, framework agreements and closed-loop data-collection deals should not automatically be presented as evidence of genuine commercial demand.
But look at the industry over a longer time horizon, and the picture becomes more complicated.
Some of these seemingly circular orders may actually represent the industry spending real money to build something it desperately needs:
the training material for the next generation of robots.
A robot can be physically built without a massive dataset.
But without sufficient real-world data to train it, it may remain little more than a machine capable of performing impressive demonstrations on an exhibition stage.
Robots cannot learn the physical world simply by waiting for better algorithms. Nor can simulation completely replace real-world experience.
They have to learn through repeated interaction with reality—one robot at a time, making mistakes, recording what happened, and turning those experiences into data.
The current system is far from perfect.
Data is not standardized.
Training environments can be artificial.
Duplicated investment is widespread.
Data remains fragmented across companies and robot platforms.
All of these problems are real.
But there is a simple way to judge whether a technology sector is genuinely moving forward:
Don’t just look at how much money it has raised. Look at whether it is creating reusable assets for the next generation of products.
That is what the humanoid robotics industry is building right now.
It is the industry’s most difficult, expensive and unglamorous “course”—but it is also one that no serious player can afford to skip.