Behind hundreds of robot training grounds, embodied intelligence begins to recalculate the accounts

+1

A year ago, one of the most popular questions in the embodied intelligence industry was still: where does robot data come from? Today, the question has quietly become: is the data we spent so much money collecting actually useful?

A recent scene at Beijing's Shougang Park is quite symbolic. According to media reports, in March 2025, Beijing's first humanoid robot data training center was unveiled here. The planned 3,000-square-meter site was to house 108 robots and build ten major scenarios including home eldercare, new retail, and automotive assembly. At the time, such training centers were seen as new infrastructure for the embodied intelligence era, and Shijingshan District even proposed that the project's future annual data output could exceed 1 million entries. Less than 18 months into operation, change came: the original operator has already withdrawn its own equipment and robots. The project side explained that the original teleoperation approach deviated from the new development direction, and the quality and precision of the collected data also fell short of expectations.

According to incomplete statistics from the "Embodied Intelligence Training Ground Research Report (2026)" (hereinafter referred to as the "Report") recently released by institutions including the China Academy of Information and Communications Technology, as of the end of June 2026, China had put more than 70 embodied intelligence training grounds into operation, with another 46 in the planning or construction stage.

At the same time, debate around data collection centers is heating up. Some industry figures question whether some data collection projects rely on government support, robot body procurement, and data buybacks to form an internal loop. Data collection centers have not disappeared, but they have entered a stage of recalculating the accounts.

The data shortage has spawned a heavy-asset business

The rapid emergence of data collection centers has a real technical backdrop: robots lack high-quality physical-world data.

Large language models can obtain massive amounts of text and images from the internet for learning. But when robots learn to fold clothes, assemble parts, or open drawers, they need to know more: at what angle the wrist turned, how much force the gripper applied, how to correct after an object slips, and how to re-complete the task after a failure. Therefore, the most direct approach over the past two years has been to build training grounds, where humans teleoperate robots to repeatedly perform operations, turning trajectories into training data.

But this model has been expensive from the start. The aforementioned Report estimates that a training ground of several thousand square meters typically requires an investment of tens of millions to over 100 million yuan. Deploying 100 humanoid robots could cost tens of millions of yuan for the robot bodies alone, followed by scenario construction, data collection equipment, computing power, networking and storage, as well as labor, maintenance, and venue rental costs.

What's more troublesome is that operating for 8 hours does not mean obtaining 8 hours of valid data. Robots fail, equipment malfunctions, scenarios need to be reset, and data still has to go through cleaning and quality inspection. Industry figures estimate that for a professional teleoperator working 8 hours a day, the truly usable data may ultimately be only two or three hours.

In the past, people were used to describing the scale of data collection centers with "100,000 hours," "500,000 hours," or "one million hours." Now it seems these numbers can at most indicate production capacity. What truly determines commercial value is how much of that data can enter models, how much can be sold, and how much smarter the models actually become after consuming this data.

How do data collection centers recoup their investment?

The first thing data collection centers think of, of course, is selling data. However, the business model of a data collection center depends on at least three variables: efficiency, sales rate, and repurchase rate. How much of the 100,000-hour capacity can pass acceptance is the first threshold; how much of the accepted data can be sold is the second threshold; and whether customers will continue buying in the second year is the third threshold.

Here there is a difficulty unique to embodied intelligence. Different robots have different degrees of freedom, grippers, sensors, and control systems, and the extent to which data collected by one robot can be transferred to another is still being explored. Therefore, the most natural customer is the company that produces that robot.

This gives rise to a second model: data buyback. For example, documents submitted by iSoftStone to the Shenzhen Stock Exchange show that its subsidiary iSoftStone Tianqi procured robots from AgiBot for building a embodied robot data collection site in Wuxi; after the project is implemented, AgiBot will conduct data buybacks as agreed, forming a closed cooperation loop of "project acquisition—implementation—data return."

The question then advances one step further: if the most important data customer of a data collection center is long-term the robot supplier itself, how large can this market expand? A robot company sells its bodies to a training ground, the training ground produces data, and the company buys it back—the two transactions generated are hard to understand as two independent market demands. This is where the current controversy over the "internal loop of data collection centers" truly deserves attention.

At present, data collection centers in some places simultaneously undertake functions of investment attraction, industrial cultivation, and infrastructure construction. According to reports, among the 64 data collection centers already in use, at least 13 have deployed more than 100 robots; many projects adopt a government-enterprise co-construction model. Some industry figures say that some projects involve arrangements in which local platforms procure robots and robot manufacturers then buy back data, and some local governments have begun to become cautious about the input-output of such projects.

The real risk may go beyond failing to recoup costs

Beyond the economic calculus, data collection centers have also encountered a more thorny technical problem: data scale and data value are not naturally proportional. This is also why the adjustment of the Beijing Shougang Park project deserves attention.

To improve efficiency, training grounds often replicate kitchens, shelves, assembly lines, and even hospital wards into a fixed space. Here lighting is stable, the floor is flat, and objects are placed relatively uniformly, so robots can repeat the same action hundreds of times a day. Real factories, however, will not cooperate in this way.

Parts may be reflective, floors may be oily, object placement varies from day to day, and production lines have their own rhythms. Once a robot enters the site, what truly determines whether it can work over the long term is often the abnormal situations that rarely appear in the training ground: failed grasps, sudden occlusions, dropped parts, equipment failures, and how to recover after a failure. Hence a seemingly counterintuitive result: a data collection center may well have produced massive amounts of data, yet the improvement in model capability is very limited.

Recently, industry practice has begun to change. Beyond the centralized "model showrooms," bodyless data collection, first-person video, simulation synthesis, and distributed real-world scene collection are all developing rapidly. Some companies put collection devices directly into homes, offices, retail, and production environments, letting data arise naturally from real activities. Other platforms choose to replicate specific workstations 1:1, having robots complete pilot testing before entering the production line, and whenever problems arise on site, they bring the failure cases back for retraining.

As a result, an important pricing change is beginning to emerge in the data industry. In the past, what was sold was "an hour of data"; in the future, customers may care more about how much the data raised the robot's task success rate, how much manual intervention it reduced, and how much deployment time it shortened. Data scale still matters, but the capability increment depends on data quality, task coverage, the proportion of hard cases, the degree of matching with the target embodiment, and the model's evaluation results.

This will make many data collection centers that tell their story through "production capacity" look awkward. A center producing 1 million hours of data a year, if the data is highly repetitive and hard to transfer across embodiments, will contribute little to model improvement, and scale itself will not automatically translate into a competitive advantage. Conversely, a platform producing only tens of thousands of hours of data, if it possesses a large number of failure cases, long-tail situations, and high-quality force feedback from real production, its data may instead be more valuable.

What the data collection industry has found easiest to measure in the past is "how much was produced"; the next stage requires answering "how much the robot actually learned."

Data collection centers will not disappear, but "data factories" might

This round of questioning does not mean embodied intelligence does not need data; high-quality physical interaction data remains scarce. The problem is that the first generation of data collection centers understood data collection too much like traditional manufacturing: build a factory, set up hundreds of machines, hire a batch of operators, and then measure output in "hours."

This model is being corrected by reality. The data collection centers that truly have vitality in the future will likely simultaneously handle data collection, model training, testing and validation, pilot testing, and the return of on-site problems. Their core assets will also change: the number of robots and the floor area will gradually recede into secondary importance, while real industry scenarios, model capabilities, and channels for continuously obtaining high-value data will become increasingly important.

This may also be the real lesson left by this round of "cooling down" in data collection centers. When a new industry first emerges, quantity is the easiest way to create prosperity: how many centers, how many robots, how many square meters of space, how many tens of thousands of hours of data—all are clear, intuitive, and suitable for being written into industrial plans and project reports.

But when the industry begins to enter the real commercialization stage, the evaluation system must change. When looking at a data collection center in the future, there are at least three numbers worth asking about: how much revenue comes from third-party customers unrelated to the embodiment supplier, whether customers make repeat purchases, and how much capability improvement the collected data actually brought to the model.

The first two numbers test whether the market truly exists; the last number determines whether this data is actually worth anything.

Over the past year, the embodied intelligence industry has been busy building "schools" for robots. Now the first batch of students has reached the school gate. Next, the standard for measuring the value of these schools will become increasingly simple: after the robot leaves the training ground, can it truly do the job well?

The Paper Weibo; The Paper WeChat official account; The Paper Douyin account; Paisheng Wanwu Open Platform; IP SHANGHAI; SIXTH TONE

Tip line: 021-962866; Tip email: news@thepaper.cn