China Mass-Produces Robot Behavior Data with Large-Scale 'Data Factories'

A clear shift in robot technology was observed at the World Artificial Intelligence Conference (WAIC) held in Shanghai last July. While robots in the past mainly demonstrated specific tasks, this event featured robots performing actual tasks such as making coffee, sorting goods, and doing laundry, placed throughout the exhibition hall. This indicates that robot technology has moved beyond the experimental stage and is now focusing on interaction and task performance capabilities in real-world environments.

Behind this transformation is the Chinese government's proactive development policy. China has designated 'Embodied AI'—where AI interacts with and learns from the real world through a physical body like a robot—as a national key technology. This means that beyond simply developing artificial intelligence, China is focusing on enabling robots to collect data in physical environments, learn autonomously, and solve real problems.

To advance embodied AI robots, China is newly promoting 'factories that teach robots.' These are factories specializing in manufacturing data to train robots, focusing on producing large-scale data so robots can learn various behaviors and interactions needed in the real world. Reporter Bae So-jin described this as "factories that manufacture data to teach robots."

This implies that infrastructure for robot training will be of national importance and systematically supported.

With the goal of developing more than 100 high-value task scenarios by the end of 2026, the plan is to enable robots to learn complex and diverse situations they may encounter in real environments. This strategy aims to expand the practicality and application range of robots.

These standards will support the efficient utilization of robot training data in major industrial fields and ensure data consistency and compatibility. This strategy aims to secure data reliability through national standards and maximize the efficiency of robot development.

For example, at the Vertical Mountain Center in Beijing, trainers work 24 hours a day in day and night shifts to train robots. This is the result of adopting an efficient operating method to continuously produce the vast data needed for robot learning.

This indicates the necessity of filtering out errors, duplicates, or unnecessary information that occur during the data collection process, and the difficulty of securing high-quality data.

To teach a robot one complete action, immense repetitive learning is essential. Depending on the difficulty of the task, hundreds to thousands of repetitions are needed, and for precise tasks, almost 10,000 data points may be required. This means trainers must demonstrate the same action repeatedly to the robot, emphasizing the importance of securing high-quality behavioral data.

These robots mimic and record human movements, thereby creating the foundational data needed for robot learning.

In conclusion, the ultimate product of robot training grounds is not a single robot skilled in a specific task. Instead, it is standardized behavioral data that can train a 'common brain' capable of controlling various types of robots. This data will be used to develop general-purpose intelligence applicable to diverse robot platforms.

Operating a single robot remotely required many personnel and time, making it inefficient for large-scale data production.

This method can secure large amounts of behavioral data much more efficiently than remote robot operation.

News reports from India describe local data companies recruiting housewives and factory workers to film their work processes, and in the US, it is becoming a new form of labor. This is a global trend to secure large amounts of real-world data needed for robot learning.

JD.com, one of China's top three e-commerce companies, is developing this data collection method into a large-scale, region-wide industry beyond individual side jobs. JD.com is primarily recruiting local residents under 50, especially stay-at-home mothers, to build a large-scale data production system. This strategy aims to both revitalize the local economy and efficiently secure vast amounts of data needed for robot learning.

It can accurately record the 3D relationship between hands and objects better than a typical smartphone, which is advantageous for acquiring large amounts of high-quality physical behavior data. JD plans to use this device for large-scale data collection projects.

China's enormous pool of young talent is being actively channeled into this data production line. This strategy leverages China's strengths—low labor costs and abundant manpower—to secure data essential for robot learning on an overwhelming scale.

In terms of labor costs, the United States finds it difficult to match China. US companies are compelled to outsource data collection to countries with lower labor costs, such as India or Africa. In contrast, China can organize large-scale data collection projects led by the government, mobilizing its domestic young technical workforce, laborers, and even local residents. This structural difference provides China with a significant advantage in the competition for robot data acquisition.

Another strength of China's robot data production system is its ability to integrate with the country's powerful manufacturing ecosystem. Chinese companies can source most of the key components for humanoid robots from their domestic supply chain. This creates an efficient virtuous cycle throughout the entire process, from component procurement to data production and the development of finished robots.

For the past two decades, China has served as the world's factory for producing goods. Now, what China aims to mass-produce is not just the physical bodies of robots. It intends to mass-produce human experience and physical intelligence to power robot bodies, as if churning them out of a factory. This demonstrates China's ambitious goal to secure 'action intelligence,' a core capability in robot technology, and seize leadership in the global robot market.