As humanoid robots danced, drummed and poured coffee at China's latest tech showcase in Fuzhou, Fujian province, one of the largest crowds gathered around something far less glamorous: a worker wearing a virtual reality headset teaching a robot how to pick up a paper cup.

For robotics company Joyful Embodied, that moment captured what its CEO said is the real battleground in the global race for embodied artificial intelligence — not flashy hardware, but data.

"We are building a school for robots," Chen Yishi, CEO of Joyful Embodied, said in an exclusive interview with China Daily, describing plans for a massive robot data collection facility in Fujian that will operate around the clock.

The company plans to build a more than 3,000-square-meter training ground packed with fleets of humanoid, quadruped and wheeled robots performing real-world tasks across household, industrial and commercial settings.

Human operators wearing VR gear and motion-control devices will remotely guide the robots through thousands of repetitive actions — stacking cups, sorting objects, carrying parts and wiping tables — while cameras and sensors capture every angle, joint movement and pressure point.

Each robot will work 24 hours a day, generating what Chen called the "high-quality fuel" needed to train the next generation of embodied AI systems.

"Without real-world data, even the most advanced large models are just empty brains," Chen said."Robots do not learn through assumptions. Every tiny movement has to be taught through data."

The effort reflects a broader shift underway in China's AI industry as companies move beyond chatbots and language models toward "embodied intelligence" — AI systems embedded in physical machines that can perceive, decide and act in real environments.

The sector gained fresh political momentum this year after embodied AI was written into China's 2026 Government Work Report as a strategic future industry. Fujian, a southeastern manufacturing hub better known for electronics and trade, is now pushing aggressively into robotics, AI infrastructure and industrial automation as part of a broader AI development drive.

While investors have poured billions into humanoid robotics hardware, industry experts increasingly argue that high-quality real-machine training data — especially teleoperation and motion-capture data — have become the sector's scarcest strategic resource.

Chen believes the economics of the industry will increasingly revolve around data rather than hardware alone.

"High-quality embodied AI training data are already priced by the hour," he said.

The company, founded in September 2025, is attempting to position itself not merely as a robot manufacturer, but as a full-stack embodied AI developer. Alongside its data operations, it is building its own large AI models and a development platform called Joyful Studio, aimed at enabling customized robot applications in manufacturing, security inspection, logistics, education and service industries.

According to Chen, every continuously operating robot inside the company's data collection system creates demand for roughly three technical positions spanning data annotation, algorithm optimization and equipment maintenance.

The company is also betting that Fujian's export-oriented economy and overseas Chinese business networks could help it expand robot deployment and data services abroad.

For now, however, the immediate race is domestic — and it's accelerating rapidly.

The Ministry of Industry and Information Technology recently approved the country's first industry standard for embodied AI benchmarking, due to take effect on Monday, marking what industry experts see as the beginning of a more formalized ecosystem for robot evaluation and deployment.

Chen said the industry is entering a decisive phase. "The competition is no longer just about whose robot can walk or dance," he said. "The real question is: whose robot can truly work."