Fujian Jufu Technology has launched a data collection facility in the Fuzhou Software Park of Fujian, China, to train humanoid robots for real-world tasks. Nearly 30 robots are currently practicing activities like sorting produce and cleaning tables under the guidance of human operators using virtual reality devices to record movement data.
The transition of artificial intelligence from digital screens to physical bodies requires a fundamental shift in how machines learn. While large language models are trained on vast archives of text, a robot cannot master the physical world through reading alone. It requires a cycle of observation, repetition, and failure. This necessity has led to the establishment of a specialized facility in Area D of the Fuzhou Software Park, where humanoid robots are undergoing a process described as going to school
.
Data Collection at the Fuzhou Software Park
The facility, created by Fujian Jufu Technology, serves as the province’s first large-scale data collection factory. Rather than relying on pre-written code to dictate every single movement, the operation utilizes nearly 30 robots that learn by mimicking human behavior. This process is overseen by operators who act as teachers
for the machines.
The training mechanism relies on a tight loop between human intent and robotic execution. Operators wear virtual reality devices and use controllers to guide the robots through specific exercises. When an operator moves their arm to pick up an object, the robot reproduces the gesture in real time. For example, a robot might grasp a paper cup and place it atop another, mirroring the exact motion of its human guide.
The objective is not merely the completion of the task, but the capture of high-fidelity data. Every movement, joint angle, and amount of pressure applied by the robot’s grippers is recorded by a network of sensors and cameras. The robots are currently practicing a set of utilitarian tasks, including cleaning tables, sorting fruits and vegetables, and discarding packaging boxes. This data collection is essential for building a library of real-world trajectories and force requirements that can be generalized across different environments.
Shifting from Manual Programming to Demonstration
For decades, industrial robotics relied on manual programming. Engineers wrote precise coordinates and logic gates to ensure a robotic arm moved from Point A to Point B with millimeter precision. This method works in a controlled factory setting where the environment never changes, but it fails in the real life
scenarios Fujian Jufu Technology is targeting.
The Fuzhou facility represents a change of paradigm: moving from manual programming to teaching through demonstration. In the physical world, variables are constant and unpredictable. A fruit may be a different shape than the previous one, or a table may be slightly tilted. By recording how a human handles these variances, the robots accumulate data on how to adjust their force and trajectory on the fly.
This approach allows the robots to learn the nuances of physical interaction—such as how much pressure to apply to a vegetable without crushing it—that are nearly impossible to hard-code. The focus is on gathering real-world movement data to bridge the gap between a laboratory prototype and a functional worker.
China’s Manufacturing Scale and Robotic Ambition
The effort in Fujian is part of a broader industrial strategy. Recent reporting indicates that Chinese humanoid robots now account for 60% of global manufacturing. This scale gives Chinese firms a distinct advantage in the iterative process of robotic development; they can deploy more hardware into training environments more quickly than competitors with smaller production capacities.
The ability to mass-produce the hardware is only half of the equation. The current bottleneck in humanoid robotics is not the agility of the limbs or the precision of the hands, but the software’s ability to operate outside of a controlled lab. The Fuzhou “school” is an attempt to solve this software problem by treating data collection as an industrial process. By scaling the number of “teachers” and robots, Fujian Jufu Technology is attempting to accelerate the creation of a dataset that can be used to train more generalized robotic brains.
The Data Bottleneck in Physical AI
The move toward demonstration-based learning highlights the primary challenge facing the robotics industry: the scarcity of high-quality, real-world interaction data. While AI models have the entire internet to learn language, there is no “internet of movement” for robots to download.
Every single physical task—from opening a door to sorting a box—must be captured manually or simulated. Simulation is fast, but it often suffers from the “sim-to-real” gap, where a robot performs perfectly in a virtual world but fails in the physical world due to unforeseen friction or gravity variances. The Fuzhou facility bypasses this by recording actual physical interactions.
The implications of this strategy extend beyond the specific tasks of cleaning and sorting. If a company can successfully build a pipeline that converts human demonstration into robotic skill at scale, the time required to deploy robots into diverse workplaces will drop significantly. The focus is no longer on building a robot that can do one thing perfectly, but on building a system that can be taught to do anything a human can demonstrate.
As these robots move from the Fuzhou Software Park into broader applications, the success of the program will be measured by how well the robots handle the unpredictability of human environments. The current phase is a test of whether the data collected from nearly 30 robots is sufficient to create a blueprint for general-purpose labor.
