For a long time, the big conversation about artificial intelligence has revolved around models capable of summarizing, programming or generating images. But when we take that ambition to the physical world, everything changes. A robot does not learn to work just by reading instructions: it needs to observe, repeat, fail and accumulate data on real movements. That is why the next frontier of robotics is not only in manufacturing more agile bodies or more precise hands, but in building the entire system necessary to teach them to act outside the laboratory.
This system is beginning to take shape in Fujian, where the province’s first large data collection factory has been launched in a test phase. According to CCTVthe facility is located in Area D of Fuzhou Software Park and has been created by Fujian Jufu Technology. There, almost 30 robots follow the instructions of different operators, described by Chinese sources as “teachers”, to practice tasks such as cleaning tables, sorting fruits and vegetables or disposing of parcel boxes.
The mechanics of that “school” are relatively easy to imagine, but very demanding underneath. Operators wear virtual reality devices and operate controls to guide the robot during each exercise. When the operator raises his arm, the machine reproduces the gesture and, for example, grab a paper cup to place it on top of another. The important thing is not only that it completes the action, but that each movement, joint angle and clamp pressure is recorded by cameras and sensors.
The school where robots learn with real data
One of the least showy parts is also one of the most decisive. The tasks we see in the video, such as cleaning a table or picking up a glass, seem simple because we do them almost without thinking. For a humanoid, on the other hand, each gesture requires a specific sequence of physical decisions. Data collection engineer Jiao Shiwei explained to Fuzhou News that even the smallest movements need to be learned through data, and that each action must be designed according to the characteristics of the robot itself to find the most suitable trajectory.
The key word here is “generalization.” That is, the ability to apply what has been learned when the environment is no longer identical to the training environment. Shiwei summed it up with two very basic actions: pick up a glass and clean a table. If the object, surface and stain do not change, the robot has it relatively easy. But in a house, a factory or a service space, almost nothing is repeated the same. Hence, data collection workers introduce variations in glasses, tablecloths and tables to expand the scope for learning.
The bottom line is that robots are also entering their own race for data. In other areas of AI, much of the progress was based on digital material already available. In robotics, on the other hand, many of the examples must be generated from scratch, with real machines, real objects and movements repeated over and over again. Xinhua puts the problem in these terms: the bottleneck of humanoids is no longer concentrated only in the hardware, but in how to continue perfecting their “brain” through training in application scenarios.
The industrial reading of the project helps to understand why these small tasks can end up becoming infrastructure. Chen Yishi, CEO of Jufu Technology, told Fuzhou News that these types of factories provide support for end-to-end models and implementation in vertical scenarios. The idea is that an AI robot does not function as a traditional machine limited to a fixed sequence, but as a guided system capable of make decisions on the body from real training.
The company is also recent. Jufu Technology was founded in September 2025 and presents its activity as a combination of data factory and self-development. Its objective is not limited to accumulating examples of movement, but to create around that base a local ecosystem of algorithmic talent, data and collaboration with the industrial chain. Yishi, for his part, pointed out that its future products aim at industrial manufacturing, safety inspection, research and education, although sources present it as a roadmap, not as an already consolidated deployment.
Images | Jufu Technology | Xinhua
GIPHY App Key not set. Please check settings