From farm work to Henan Opera: Inside a 'preschool' for humanoid robots

It is hard to imagine, without seeing it up close, a humanoid robot with a rigid metallic frame singing an excerpt from the classic Henan Opera Hua Mulan with controlled phrasing and stylized movements, or beating a drum and dancing to the rhythm of the popular song Da Hua Jiao (Wedding Sedan Chair).

But a recent visit to Henan Embodied Intelligence Industry Development Co, Ltd in Zhengzhou, capital of Henan province, challenged my long-held impression of cutting-edge technology as cold and distant.

The robots on display were strikingly versatile. In simulated farming scenarios, they performed agricultural tasks with well-practiced movements. In industrial settings, they handled factory logistics with precision. A moment later, the same machines could pick up a calligraphy brush, practice tai chi with fluid motions, or serve as exhibition guides.

How did these robots acquire such a wide range of skills across agriculture, industry, culture, and entertainment — and how are those skills being brought closer to real-world use?

As China enters the 15th Five-Year Plan (2026-30) period, the country has laid out a clear direction for intelligent industries: developing new quality productive forces, driving industrial innovation through technological innovation, advancing the "AI Plus" initiative, building application scenarios for future industries, and speeding up the transformation of frontier technologies into real-world productivity.

Embodied AI and humanoid robots are seen as important drivers of this process. The key question is how to bridge the gap between research and real-world application so that advanced technologies can truly serve industries and households. This training ground in Henan offers one practical answer.

Chen Yang, business manager at Henan Embodied Intelligence Industry Development Co, Ltd, used a vivid analogy to explain the process.

"When a humanoid robot first comes off the production line, it is much like a newborn baby. Its understanding of the physical world is almost blank," Chen said. "Our training center is like a preschool, and our researchers are like parents teaching the robots different skills."

Behind the simple analogy is a demanding process of data training. Chen told me that even stabilizing a basic movement can require hundreds or thousands of simulations. To build a scalable and generalizable model, tens of thousands or even hundreds of thousands of training iterations may be needed.