What does 22,000 units mean? Shanghai has about 25 million residents; if you moved all these robots there, fewer than one would be available per 10,000 people. On October 16, 2026, at a meeting of the ITU Focus Group on Embodied Intelligence Standardization in Hangzhou, the China Academy of Information and Communications Technology (CAICT), together with Tsinghua University's Department of Electronic Engineering and an embodied intelligence testing lab, released the English edition of the *Embodied Intelligence Development Report (2026)*. Its figure: global humanoid robot shipments exceeded 22,000 units in the first half of the year, up nearly 300% year-over-year.
The same day, in Guangzhou, in the service robotics zone of the 140th Canton Fair, media cited IDC data: global shipments approached 25,000 units in the first half, up 432.1% year-over-year. Put the two numbers side by side and a gap emerges: about 13% difference in units, and 132 percentage points difference in growth rate. In the same half-year, two authoritative methodologies described an industry with completely different growth trajectories.
(The dashed gap between the two figures reflects unexplained differences: statistical scope, morphology definitions, and sample sources were not clarified in public reporting. This article takes no side.)
The Structure: Where Do 100 Robots Go?
Shipment figures can be argued over, but the distribution across scenarios reveals the industry's real situation. The CAICT report divides first-half humanoid deployments into four buckets: entertainment/commercial performance plus science education and data collection, over 60%; guided tours and retail assistance, about 19%; smart manufacturing, about 13%; warehouse logistics, about 5%. Out of every 100 robots, roughly 60 earn their living at exhibitions, on stages, and in classrooms (with a batch in data collection centers); about 19 greet customers at malls and bank entrances; only about 13 actually go onto production lines and about 5 into warehouses.
This distribution reads differently against September's "robots entering factories" narrative. On October 16, Leju Robotics and Zhaofeng Co. widely publicized their automotive parts loading solution in Zhejiang: 15 humanoid robots running in normal operation on a production line, single cycles under 40 seconds, covering 15 workstations and 10+ part types, with a stable 98% success rate and about 1,000 cumulative hours of operation. The individual case is solid. But placed back into the 13% structure, its identity is clear: manufacturing is a direction that is growing—not yet the industry's bread and butter.
Over a longer horizon, the trend is visible. When we covered the IDC report in late September, display-oriented scenarios had fallen from 83.8% a year earlier to 69%. Now CAICT's combined entertainment and education/data-collection figure still exceeds 60%. The classification systems differ and cannot be directly spliced, but the direction agrees: robots sustained by performance and teaching still account for close to half.
Industry Chain and Geography: Concentrated Head, Long Tail
The report describes a three-layer industry chain: upstream for data and compute, basic software/hardware, and core components; midstream for embodied intelligence products and technical services; downstream for industry applications. Geographic concentration is higher: Beijing hosts 109 listed companies, Guangdong 84, Shanghai 58—together about 62%. Jiangsu and Zhejiang have 45 and 39 respectively; the top five provinces/cities exceed 80%. Beijing concentrates on algorithms, models, and complete robots plus data collection equipment; Guangdong and Shanghai lean toward complete robot R&D and basic components; Jiangsu and Zhejiang cluster in product support and industry applications. The report's phrase: "concentrated head, long tail."
Official Tone: From Technical Closed-Loop to Value-Driven
Wang Yu, professor of electronic engineering at Tsinghua University, made remarks at the release worth reading closely: embodied intelligence is at a critical window moving from a "technical closed-loop" toward "ecosystem synergy"; the industry's development will not be completed by a single technical breakthrough, but more likely through spiral, incremental evolution. Technically, the report argues AI's underlying architecture needs rethinking so that models, the embodiment, and the environment form tighter dynamic coupling. Industrially, the shift should be from "technology-driven" to "value-driven," doing deep and solid work within capability boundaries.
In plain terms: don't wait for miracles, and don't try to do everything. Deepen reachable scenarios first, get models, robot bodies, and the environment genuinely working together—solo score-chasing won't go far. This judgment aligns well with the data: 60% entertainment/education, 13% manufacturing.
Two Numbers Clash—Which to Trust?
Both institutions are authoritative, yet the numbers don't match. Possible explanations: different statistical scope (whether wheeled and semi-humanoid forms are included), differences between self-reported and channel-verified data, and inconsistent definitions of "humanoid robot." These are speculation—neither report detailed its methodology, and the full English report was not retrievable from public channels. One thing is certain: when the industry starts seriously debating methodology, it has moved past the "just report a number" stage.
The structure is what really matters. Over 60% in entertainment and education/data collection means today's main paying customers for humanoid robots are exhibitions, education, and research data—productivity is still outside the door. Manufacturing's 13% is the key stake in the next narrative. Leju and Zhaofeng's 15-unit production line—and their donation of real operational data to the OpenAtom OpenLET community—is exactly brick-laying into that 13%. If the mid-year report next year shows manufacturing plus warehousing/logistics rising from 18% to over 30%, "robots entering factories" will have moved from isolated cases to an industry.
Sources and caveats: CAICT figures come from a Xinhua-owned China News Service dispatch from Hangzhou (Oct 16) as relayed by *Science and Technology Review*; the full English report was not obtained. IDC figures come from a China News Service Canton Fair report; the original IDC document was not verified. Leju–Zhaofeng project details come from a China Economic News Network report. The IDC display-scenario share (83.8% to 69%) follows this publication's September 30 article. The "22,000 vs. 25,000, ~13% difference" and "132-percentage-point growth gap" are our own calculations from the two disclosed figures. Wang Yu's remarks are relayed from the release event; no full transcript was seen.