Inside Shenzhen: How China Trains Humanoid Robots With Human Bodies and Millions of Hours of Real-World Data

Shenzhen pulses with the clatter of factories. Here, in China’s hardware capital, a new workforce has emerged. Workers strap on VR headsets and motion-tracking gear. They guide humanoid robots through tasks once reserved for human hands. The goal isn’t just remote operation. It’s data. Vast troves of movement data to teach machines to act on their own.

At IO-AI Tech, a startup 45 minutes north of downtown, this process unfolds daily. Employees don headsets reminiscent of Ready Player One. They control robots in mock convenience stores and factory lines. One moment a worker stacks boxes. The next, a robot mirrors the motion. The system captures every adjustment, every hesitation. That information feeds AI models hungry for physical understanding.

Teleoperation as Training Ground

Will Knight visited the facility for WIRED. He tried the gear himself. “I’m a little embarrassed to say that the first thing I tried with this futuristic gear was getting all 10 hands to flip the bird,” he wrote. After that, he stacked shelves. “It was disorienting at first: I had to adjust to a slight difference between my movements and those of the robot I could see through the headset. After a little practice, however, I was stacking shelves like a robot-boss.”

Si Chin, cofounder of IO-AI Tech, chose Shenzhen for a reason. The city houses thousands of manufacturers. Prototypes move from idea to test in days. The company works with local firms such as Jack Sewing Machines. Robots learn to iron shirts on production lines. They fold laundry in simulated apartments. One demo showed a Unitree humanoid. A human operator marched beside it. The machine copied steps exactly. Then it reached for a shirt on a hanger, removed it, and folded the garment. All while the operator viewed the scene through the robot’s cameras.

But. The real work happens off the demo floor. Operators control multiple robotic hands at once. One system tracked 10 hands from different makers. A custom glove transferred finger movements to 50 robotic digits instantly. Feedback flowed back. The operator could feel a ball placed in an electronic palm. Bidirectional. Immediate. This isn’t flashy theater. It’s deliberate data collection aimed at specific problems.

Si Chin explained the approach to Knight. The company wants robots stocking shelves and picking items from bins today. Tomorrow, that data trains autonomy. “It is similar to self-driving cars,” she said. “You need this training data that’s more focused on the specific thing you’re trying to address.” Some roboticists believe vast teleoperation datasets will unlock general models. Capable ones. Adaptable across environments.

And the scale is growing. In homes across China, people invite robots in. Shenzhen-based X Square Robot sends units to apartments. They perform chores slowly. A robot might fold three pieces of clothing in an hour. A human housekeeper handles most of the work. The machine’s cameras record everything. Daniel Wang, in Beijing, tried the service. He paid 149 yuan, about $22, for three hours. “In terms of privacy, I’m okay with showing these [home] scenes,” he told Rest of World. “I feel like I made some contribution to physical AI.”

Elsewhere, stay-at-home moms earn extra. Gao Bo, in her 50s in Shandong, films chores six hours a day. She earns 20 yuan an hour. “No one had paid me to cook and do laundry before,” she said. Her apartment ended up spotless. JD.com pursues even larger numbers. The e-commerce giant recruits 100,000 employees and 500,000 external workers in one city alone. Target: 10 million hours of egocentric video over two years. Head-mounted cameras. Wrist sensors. Factories in Guangdong. Elderly care centers. Kiwifruit farms. All feed the same pipeline.

This human labor creates an edge. Chinese firms gather data at low cost. Workers eager for side income. Government support. Public enthusiasm. Compare that to higher expenses elsewhere. The result shows. China produced 12,800 humanoids in 2025. Roughly 90% of the global total, according to a MERICS report. Many landed in training centers, labs, logistics, and manufacturing. Output could nearly double in 2026. Projections point to aggressive scaling.

UBTech in Shenzhen delivers hundreds of Walker humanoids already. Plans call for thousands more this year. Some head to facilities near the Vietnam border. EngineAI opened a plant that assembles one robot every 15 minutes. Unitree ships thousands. Its G1 model appeared in high-profile demos. LimX Dynamics raised $200 million. X Square Robot secured 1 billion yuan. Capital flows. Yet investors grow picky. Flashy videos no longer impress. They demand robots that operate in factories for months. That cut costs. 2026 marks an elimination round for many of the more than 200 embodied AI startups.

Challenges remain. Robots still stumble. Balance fails in unfamiliar settings. Teleoperated moves don’t always translate when shapes, sizes, or weights differ. Autonomy must bridge those gaps. Data helps. But volume alone doesn’t guarantee success. Models need curation. Tasks must match real deployment. Chinese officials push hard. A national initiative targets 10,000 commercial humanoid deployments by year’s end. Over 100 high-value applications in manufacturing, healthcare, warehouses, and disaster relief. “Humanoid Robot-as-a-Service” models lower barriers. Companies pay for performance, not hardware.

Shenzhen leads the charge. Its supply chain produces actuators, sensors, and motors cheaply. A full humanoid can cost $10,000 to $30,000. Far below many Western counterparts. Local governments fund testing centers. Vocational schools explore teleoperation as a trade. Young trainers, many born after 2000, guide robots through simulated scenarios at centers like the one in Hubei. They spend hours on everyday tasks. The machines learn nuance. The humans earn wages in a shifting economy.

Production ramps. Data accumulates. Deployments spread. Factories test robots on assembly. Stores trial shelf stocking. Homes experiment with assistance. Success isn’t assured. Many units still require close supervision. Buyers hesitate without clear returns. But the infrastructure builds fast. Training centers coordinate datasets across Beijing, Shanghai, and Hefei. Innovation hubs test applications. Shenzhen’s embodied intelligence demonstration area prepares for open-street training after closed sessions.

So the workers keep moving. They fold shirts. Stack boxes. Iron fabric. Each repetition captured. Each adjustment logged. Those hours compound. They teach robots not just to copy but to adapt. China bets the physical world yields to persistence and scale. Observers watch closely. The hardware capital doesn’t wait for perfect autonomy. It trains toward it. One human-guided motion at a time.