Summary: On October 12, 2026, Shanghai launched a smart manufacturing program targeting 8,000 humanoid robots in factories by 2028, unveiled at the 26th China International Industry Fair. Siemens China’s research head pushed back the same day, arguing no factory wants its production line to be an AI testing ground. The gap between policy targets and deployment reality is the real story for anyone building in industrial robotics.
The target, and the pushback
Shanghai’s plan, reported by TechOrange citing The Paper, includes four major programs and 19 key tasks. By 2028 the city wants industrial robot density of 600 per 10,000 workers, 500 advanced smart factories, and 300 AI agents in production. Subsidies of up to 10 million RMB already exist for embodied-AI manufacturing use cases.
Zhu Xiaoxun, head of Siemens China Research Institute, was blunt at the same event: no factory is willing to turn its production line into a trial-and-error field for AI. That is the structural tension — robots learn through iteration and post-training on data, but industrial settings tolerate zero errors.
Siemens’ answer: fail in the digital twin
Siemens does not build humanoid robots. Wu Zhili, who leads equipment-manufacturing sales for Siemens’ digital industrial automation in Greater China, describes the company as the one helping robot makers design, assemble, and integrate — then bringing those robots into Siemens’ own plants.
The method is digital twins: robots iterate repeatedly in virtual replicas of real equipment before touching a live line. Siemens applies the same approach to manufacturing robot parts. Planetary roller screws — 16 to 20 per humanoid on some technical routes, each taking 2 to 4 hours to machine, per IT Times — are a production bottleneck, and Siemens simulates machining paths virtually before committing to production.
Even so, deployment is early. Zhou Yurong, GM of Siemens Numerical Control Nanjing, said humanoid robots now handle automated material feeding at the start and end of an SMT line — still a proof of concept. His words: “The Nanjing plant is not a fully automated factory.” Small batches and many product types make even line replenishment a complex planning problem.
The first real jobs are boring ones
Not everyone thinks humanoids are the answer. Zhou Bin, CEO of Shanghai Runke, argues humanlike design fits human living spaces, not complex industrial work, and showed a non-humanoid rover robot for inspection and emergency response.
Where humanoids do land, the form factor varies. JAKA says its wheeled Kargo robot is in volume production at a major auto-parts supplier’s headlight plant, cutting per-box cycle time from 105 to 28 seconds and shrinking deployment from 12 months to 57 days. Sainard’s container-loading robots target a simpler truth: many loading workers are over fifty, container interiors exceed 50°C in summer, and younger workers are scarce. As the industry saying goes, “logistics work doesn’t need legs.”
For builders, the takeaway echoes what I noted in coverage of Agility’s humanoid training facility: the winners here are the ones who pick narrow, painful, measurable jobs rather than chasing a general-purpose demo.
Standards before scale
Zhu Xiaoxun named a concrete pain point from embodied-AI vendors: getting autonomous robots into traditional automation systems. China released national standard GB/T 47859-2026 on July 2, 2026, effective February 1, 2027, to unify information models across devices and platforms and ease IT/OT integration. Siemens (China) helped draft it, and Zhu sees it as a shared language that lets autonomous and legacy equipment communicate.
The industry is also splitting into three layers. Schaeffler’s Sun Haitian, citing about a year of practice, describes distinct segments for models, robot bodies, and deployment applications.
What has to be true by 2028
IT Times frames industrial rules as zero tolerance, replicability, and clear economics. By that measure, 8,000 humanoids by 2028 is a policy number, not a deployment forecast: the Nanjing pilot is still POC, and the enabling standard lands only in 2027. If you’re building for this market, two questions matter more than the headline target — can your robot prove stability on a real line, and does the factory have the digital foundation to let it in?