A humanoid robot picks up a tray. It fumbles. It tries again. And again — up to 600 times in a single day, while sensors log every degree of joint rotation and every millisecond of grip pressure. This isn't a lab demo. It's a class schedule.
In July 2026, China opened what press coverage is calling its first "school for robots": national-level vocational training academies where humanoid robots — not humans — are the students. The goal is blunt: turn showroom robots into employable ones, at industrial scale, before global competitors catch up. Here's what's actually happening inside these facilities, and why it matters for anyone tracking where physical AI is headed next.
TL;DR
| Question | Answer |
|---|---|
| Where is the "robot school"? | Two flagship sites: 5,000 sq m in Shanghai's Zhangjiang high-tech zone, and a national vocational center in Hangzhou, Zhejiang |
| Who runs it? | The National and Local Co-built Humanoid Robotics Innovation Center |
| How many robots? | 100+ models (Shanghai) from ~12 companies; 140+ robots (Hangzhou) from ~40 firms and universities |
| What do they train on? | ~45 "atomic skills" (grasp, pick, place, transport) across ~10 task categories: domestic, industrial, medical, tourism, agriculture |
| How much data per day? | ~50,000 data points/day at Shanghai — about 10 million a year |
| Is this the only facility? | No — China has 40+ state-backed robot data centers nationwide, 24 already operational as of late 2025 |
| Is this like India's camera-training program? | Different method, same target — see comparison below |
| What's the end goal? | A general-purpose "super brain" model trained across all participating robots and manufacturers |
What China's robot schools actually are
The headline site is in Shanghai's Zhangjiang district — a 5,000-square-meter facility operated by the National and Local Co-built Humanoid Robotics Innovation Center. More than 100 humanoid models from roughly a dozen companies train there side by side, deliberately mixing hardware architectures so researchers can compare how differently built robots perform the same task (New Atlas).
A second, larger site functions as the world's first national-level vocational training facility for humanoid robots: located in Hangzhou, Zhejiang Province, it hosts 140+ robots from nearly 40 technology firms and academic institutions, running them through 40 simulated environments modeled on real workplaces (CNBC; Al Jazeera).
General manager Xu Bin described the mission as enabling "large-scale data sharing and utilization" across manufacturers — pooling training data instead of letting each robotics company hoard its own, which is a meaningfully different model than how most Western humanoid labs (Figure, Tesla, 1X) currently operate in isolation.
What "curriculum" means for a robot
Forget lesson plans. The training breaks physical labor into roughly 45 atomic skills — grasping, picking, placing, transporting — the building blocks every higher-level task is composed from. A robot might repeat a single wrist rotation or grip adjustment up to 600 times a day while sensors capture the outcome of each attempt.
Ten priority task categories anchor the curriculum, spanning:
- Domestic labor — folding clothes, tidying, retrieving items from shelves
- Industrial assembly — the same repetitive, standardized motions that made robotic arms viable on car lines decades ago
- Medical support
- Tourism and service
- Agriculture
Director Yang Zhengye's stated goal is for robots to eventually "accomplish more complex tasks requiring a sequence of actions based on their autonomous decision-making" — chaining atomic skills into full jobs rather than executing pre-scripted routines.
This is a national strategy, not one building
Shanghai and Hangzhou are the visible flagships, but they sit inside a much larger buildout. China had established more than 40 state-backed robot data collection centers by the end of 2025, with 24 already operational, spread across provinces including Anhui, Zhejiang, and Shandong (Interesting Engineering).
One standout: a 10,000-square-meter facility in Shijiazhuang, Hebei, operated by Leju with state backing, running 16 separate humanoid training programs. It generates roughly 6 million data entries a year — reportedly the highest of any single Chinese site — and its trained robots have acquired 20+ functions with a 95% task success rate.
The commercial payoff is already showing up in sales figures: data-collection centers in Jiangxi, Guangxi, and Shandong contributed to roughly 566 million yuan (about $79 million USD) in humanoid robot sales for UBTECH Robotics alone.
Deep Robotics CTO Li Chao summarized the logic driving all of this: "Unless you train robots in real situations, you cannot truly advance the technology." Simulation gets a robot most of the way there; only messy, real-world friction — bad lighting, human interference, worn tools, unpredictable surfaces — exposes the failure modes that matter for actual deployment, and government-subsidized adoption programs are what let manufacturers discover those failure modes at scale.
The end goal: a "super brain"
The stated long-term ambition at the Shanghai center is a general-purpose foundation model — informally described as a "super brain" — trained on aggregated data from every participating robot and manufacturer. That's the same idea behind vision-language-action (VLA) models like the ones covered in explainx.ai's report on Xiaomi-Robotics-1's 100K-hour UMI pre-training: the more diverse, real-world motion data you pool, the faster a foundation model generalizes to a new robot body or a new task it has never explicitly seen.
Two paths to the same data problem: China vs. India
China's academies aren't the only aggressive answer to the "how do robots learn physical tasks" problem. explainx.ai has also covered how Indian workers wear head-mounted cameras to record themselves folding laundry and working factory lines, generating egocentric footage that trains the same category of Large Behaviour Models before any humanoid hardware gets involved.
| China's robot academies | India's camera-worker programs | |
|---|---|---|
| Who performs the task | The robots themselves | Human workers, on camera |
| Funding model | State-backed, centralized facilities | Venture-backed startups (Human Archive, Objectways, Egolab.AI) |
| Data source | Robot joint/grip telemetry, direct | Egocentric human video, transferred to robots |
| Labor cost | None during collection — robots don't get paid | |
| Scale signal | 40+ state centers, millions of data points/year | Thousands of workers across garment and domestic sectors |
Both strategies point at the same bottleneck: humanoid robots can't learn to move through the physical world from text or simulation alone, echoing the argument Yann LeCun has made about why LLMs alone can't handle physical agency. Someone — human or robot — has to generate the motion data first.
Why this matters beyond China
For explainx.ai's readers tracking where AI skills are headed, three signals stand out here:
- Physical AI is now an industrial policy priority, not a lab curiosity. Dozens of state-backed facilities generating millions of labeled motion samples a year is infrastructure spending on the scale of semiconductor fabs, not a research grant.
- Deployment targets are concrete and near-term. Manufacturing and logistics — repetitive, standardized tasks — are the named first markets, the same pattern seen with Figure AI's robots already outnumbering employees at its BotQ factory.
- Data-sharing across manufacturers is the real competitive edge. Pooling training data from ~40 companies into one national dataset compounds faster than any single company's in-house pipeline — a structural advantage that's hard for fragmented, IP-guarded Western robotics labs to replicate quickly.
What people are asking
Are these robots actually being deployed after training, or is this just research? Both. UBTECH's ~566 million yuan in robot sales tied to these data-collection programs shows real commercial deployment is already happening alongside ongoing training — it's not purely a research exercise.
Does this compete with US humanoid robotics labs like Figure or Tesla? Directly, yes, on timeline and data volume. China's centralized, state-subsidized, cross-manufacturer data-sharing model is structurally different from the more siloed approach at most US and European humanoid startups, and the sheer number of state-backed sites (40+) suggests a faster aggregate data-generation rate.
What happens to the humans whose jobs these robots are training for? The stated priority sectors — industrial assembly, domestic labor, medical support, tourism, agriculture — overlap heavily with jobs held by low-wage and informal workers today. explainx.ai's coverage of India's job automation risk and the broader AI jobs data check go deeper on how fast this transition is actually moving versus the headlines.
Related reading
- Indian Workers Are Wearing Cameras to Train Humanoid Robots
- Figure AI: Robots Now Outnumber Humans at the Company
- Xiaomi-Robotics-1: 100K Hours UMI Pre-Training for Robot VLAs
- Can LLMs Drive Cars or Plan in the Real World? Yann LeCun vs the AV Optimists
- Did AI Take Jobs? A 2026 Data Check
- How to Survive the AI Apocalypse: A Job Automation Guide
Official coverage: New Atlas · CNBC · Interesting Engineering · Global Times
Facility counts, robot enrollment figures, and data-volume statistics reflect reporting available as of August 8, 2026. China's robot training network is expanding quickly — check the linked sources for the latest counts.