In a sprawling 12,000 square meter facility in Wuhan, young Chinese graduates spend their days teaching humanoid robots to serve steamed buns, wipe tables, and fold laundry. Every movement these machines make is tracked by cameras and sensors, creating what officials hope will become a vast pool of training data to power China’s ambitions in embodied intelligence. But as the country pours billions into robot training farms, fundamental questions about data transferability and technological effectiveness remain unanswered.
The Data Collection Challenge
“We’re like teachers and the robots are our students,” said Zhang Jia, a 21-year-old program manager at the Hubei Humanoid Robot Innovation Center. “When you teach a human, they get it after a few repetitions. But teaching a robot is different – you have to repeat actions hundreds, thousands, even tens of thousands of times.” This painstaking process produces about 100 hours of usable data daily, with staffers labeling every few seconds of footage with annotations like “pivot left” or “extending arm.”
The push is part of President Xi Jinping’s drive to make China a science and technology superpower. Beijing recently identified “embodied intelligence” as one of six future industries to foster in its 2026-30 five-year plan, calling for training centers, AI models, and hardware to accelerate humanoid robot deployment. Local governments from Hangzhou to Mianyang are pouring money into new training farms, with Hubei province unveiling a Rmb10 billion state fund for humanoids.
The Transferability Problem
Experts warn that a more fundamental challenge looms over China’s ambitions: data collected from one robot cannot easily power another with different hardware. “Because the hardware is rapidly evolving, data collected now may not work as well for next year’s model,” explains the primary source. This creates a moving target for researchers trying to build generalizable AI models for robotics.
Jay Huang, head of Asia industrial technologies at Bernstein research group, notes that data transferability remains an area of active research, though progress is expected. “Government support means the data is shared, benefiting everyone,” Huang said. “It pushes everyone to work in the same direction.”
Western Approaches to the Same Problem
While China focuses on physical training farms, Western companies are taking different approaches to the same data challenge. ABB Robotics and Nvidia have partnered to create RobotStudio HyperReality, which uses physically realistic simulations with up to 99% accuracy to train industrial robots. “Instead of needing thousands of physical test runs, prototype and expensive parts, robots can see and learn and understand inside a simulation that then translates perfectly into the real world,” said Marc Segura, ABB president.
This simulation-first approach claims to reduce costs by up to 40%, cut setup times, and accelerate time-to-market by 50% by eliminating physical prototypes. Foxconn is already piloting the program, with commercial release scheduled for the second half of 2026.
The Funding Race for Physical AI
Meanwhile, investment in AI that understands the physical world is reaching unprecedented levels. Yann LeCun, Meta’s former chief AI scientist and Turing Award winner, recently raised $1.03 billion for his startup Advanced Machine Intelligence Labs – Europe’s largest seed funding round. The company focuses on developing “world models” that understand physical environments through videos and spatial data.
“Anything that involves understanding the real world, we think large language models, and generative AI in general, is not the right solution,” said Alexandre LeBrun, the company’s CEO. The startup’s $3.5 billion pre-money valuation reflects investor confidence that physical AI represents the next frontier beyond today’s language models.
Practical Applications vs. Humanoid Hype
Not everyone is convinced that humanoid robots represent the most practical path forward. Mind Robotics, an industrial robotics lab spun out from electric vehicle maker Rivian, recently raised $500 million to develop traditional factory robots rather than humanoids. “Doing cartwheels does not create value in manufacturing,” said RJ Scaringe, Rivian’s CEO and chairman of Mind Robotics.
The company plans to use data from Rivian’s factory to train industrial robots for more dexterous and adaptable tasks, focusing on immediate practical applications rather than humanoid forms. This approach highlights a growing divide in robotics strategy between humanoid-focused companies and those prioritizing functional efficiency over human-like design.
The Bottom Line for Businesses
For companies considering robotics adoption, several key takeaways emerge. First, the data challenge remains significant regardless of approach – whether through physical training farms like China’s or simulation platforms like ABB-Nvidia’s. Second, hardware compatibility issues mean early investments in robot training may have limited shelf life as technology evolves. Third, the choice between humanoid and traditional robot designs depends heavily on specific use cases and cost-benefit analyses.
As Bernstein analysts note, robot purchases by data collection centers have helped sustain China’s humanoid robot manufacturers while real-world demand for their hardware is still emerging. Sales for data collection made up about one-fifth of China’s more than 20,000 humanoid robot shipments last year. But at one data collection center visited by reporters, a dozen humanlike robots hung motionless on one side of a grand lobby – not for collecting training data, but “to perform when officials come to visit.”
The race to solve robotics’ data problem is accelerating, with billions flowing into competing approaches. Which strategy will prove most effective remains to be seen, but one thing is clear: the company that cracks the code for scalable, transferable robot training data will have a significant advantage in the coming automation revolution.