Embodied Intelligence Daily · October 25, 2026

Today's Highlights

• Zhiyuan (Agibot) put 110 Expedition A3 humanoid robots dancing in sync in Jinan, setting the Guinness World Record for "most humanoid robots dancing simultaneously" — mass-production units, fleet scheduling, centimeter-level positioning
• DeepRobotics unveiled five results at once: world model, VLA, agent, policy optimization, and data closed-loop — completing its full "embodied brain" stack
• CLS (Cailian Press) reported this morning: robotics IPO thresholds have risen, shifting primary-market expectations for embodied intelligence
• The data-collection business is under scrutiny: orders are collectively shedding inflated figures, and star companies' valuation bubbles are being punctured
• Elon Musk opened a "school" for Optimus: build 10,000–20,000 robots first and let them produce their own training data

Detailed Content

1. 110 humanoid robots dance in sync; Guinness official certifies on site

• Sources: Jiemian News / DoNews / 10jqka
• Link: https://www.donews.com/news/detail/8/6752419.html
• Date: 2026-10-24

Zhiyuan organized 110 full-size Expedition A3 humanoid robots dancing simultaneously at the Yellow River International Convention and Exhibition Center in Jinan, Shandong, setting the Guinness World Record for "most humanoid robots dancing simultaneously," with an official adjudicator present to verify and issue the certificate; the confirmed successful count was 110. The robots followed unified commands, achieving high-precision synchronization in dance, positioning, jumps, and flips, using Zhiyuan's LinkBots fleet scheduling solution. Three keywords from media coverage stand out: mass-production models, fleet scheduling, and centimeter-level positioning — meaning the performers were shipping-state products, not specially tuned prototypes. Zhiyuan also revealed its robots are piloting at Jinan's Olympic Sports Center and Zhuanshanxi Road metro stations. If Saturday's Shanghai competition used black-box blind testing to make robots "actually work," Jinan's group dance demonstrated the other side: engineering capability in fleet scheduling and positioning at scale.

2. DeepRobotics releases five results at once: from world model to data closed-loop

• Sources: China Securities Journal / Mydrivers / Economic Observer
• Link: https://news.mydrivers.com/1/1160/1160221.htm
• Date: 2026-10-23/24

DeepRobotics (Yunshenchu) published multiple papers around its in-house embodied brain, DEEPRobotics-Brain, announcing five research results: world model DR-WM-1.0 (with physical constraints), embodied foundation model DR-VLA-2.0, embodied agent DR-Agent, policy optimization framework DR-APO, and data closed-loop platform DR-Benchmark. The company says the five results cover the full embodied-brain technology chain as coordinated work. A company that started with quadruped and wheeled-leg robots has laid out the entire chain — perception/prediction, foundation model, agent, optimization, data feedback — in one move. For context: per a Caixin September report, DeepRobotics ranked fourth globally in embodied-intelligence robot revenue in 2025.

3. CLS morning report: robotics IPO thresholds raised; primary-market expectations shift

• Source: Cailian Press (cls.cn)
• Link: https://www.cls.cn/detail/2488487 (September 21 verification report, for context)
• Date: 2026-10-25 morning

Cailian Press published "Robotics IPO thresholds raised; embodied intelligence primary-market expectations have changed." This is the latest link in a tightening thread since September: in late September, CLS verified with investment bankers rumors of "window guidance raising humanoid robot IPO review thresholds," focusing on revenue authenticity, profitability stability, and commercialization capability; CNBC reported in late September that regulators plan three criteria; Unitree's post-IPO share-price rollercoaster is seen as one trigger. In mid-October, 21st Century Business Herald reported that Zhiyuan, valued at 50 billion yuan, may fail to complete an IPO this year. Between "verifying rumors" and "expectations changed" in the headline lies a chain of events — exposed related-party transactions at data-collection centers, orders shedding water. Investors' acceptance criteria for robotics companies are shifting from imagination to cash collection.

4. Data-collection turmoil continues: orders deflate, the data business under pressure

• Sources: Huxiu / Sohu Tech
• Link: https://m.sohu.com/a/1092093264_116132
• Date: 2026-10-24

Huxiu published "Will data collection become embodied intelligence's sacrifice?" Core judgment: humanoid robot orders are collectively shedding inflated figures, star companies' valuation bubbles are being punctured, and if the market turns turbulent, the data-collection business will get harder. This scrutiny began in September: Mecamander founder Shao Tianlan publicly attacked "matchmaking-style entrepreneurship," accusing some high-profile, high-valuation embodied intelligence companies of using data-collection centers' related-party transactions with local governments, investors, and suppliers to manufacture unsustainable revenue for IPO sprints; multiple outlets confirmed through field research that many government-enterprise data-collection centers nationwide rely on fiscal subsidies, their core function being to inflate company revenue and package valuations. Data collection was supposed to be cold-start data infrastructure; it is now the first examination room for revenue authenticity — two sides of the same coin as the IPO threshold story above.

5. GDPS competition day two: unscripted contests push the "hidden players" forward

• Sources: Shangguan News / Jiefang Daily
• Link: https://finance.sina.com.cn/roll/2026-10-24/doc-iniwiutr7274199.shtml
• Date: 2026-10-24

The 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition entered competition days. Beyond nearly a thousand teams, nearly 200 supply-chain and supporting companies gathered — Shangguan noted a structural fact: joint modules alone consume nearly half of a humanoid robot's cost, so component vendors reach the spotlight before integrators. Three samples: Huachuang Zhiqu brought a robot with its own few-tooth-difference reducers, claiming stiffness, end-load capacity, and lifespan exceed harmonic reducers by 5x+; Banxing Technology, a low-profile ODM, drew crowds with its 800 N·m peak-torque planetary joint motors, 1000 Hz electronics, and in-motor water cooling — one of the few ODMs to have completed a robot marathon, serving 20-30 B-end customers, mostly automotive; Linyan Medical showed AR spine-surgery navigation, having completed in August the world's first humanoid-robot spine laminectomy animal experiment (pig), letting robots learn from the surgeon's first-person view and absorb radiation instead of doctors. Reposting outlets ask bluntly: of 30,000 embodied intelligence companies, whose robots actually work?

6. Musk opens a "school" for Optimus: build 10,000–20,000 robots to generate their own data

• Sources: Sina Finance / Business Insider
• Link: https://finance.sina.com.cn/wm/2026-10-24/doc-iniwhnxy1626297.shtml
• Date: Q3 2026 earnings call (10-21; Chinese coverage 10-24)

At Tesla's Q3 2026 earnings call, Musk disclosed a robot training system called "Optimus Academy," with plans to build roughly 10,000–20,000 robots that generate training data through autonomous practice. He likened it to raising children: first observe humans doing tasks, then practice repeatedly. Why a school? Tesla's autonomous driving improves via data returned from millions of cars on the road; humanoid robots lack that channel — seeing a cup isn't knowing how to pick it up; grasp position, force, and safe handling require massive practice data. So build robots first, and let robots serve as textbooks for robots. This is the most blunt version of "data infrastructure first": the model waits for data, data waits for machines, machines wait for the factory.

7. Wang Xingxing reiterates to China Daily: ChatGPT moment within one to two years

• Source: China Daily
• Link: https://www.chinadaily.com.cn/a/202610/23/WS6adab899e4b06d4aa0564592.html
• Date: 2026-10-23 (China Daily follow-up 10-25)

Wang Xingxing (Unitree) told China Daily that the "ChatGPT moment" for humanoid robot AI could arrive within as little as one to two years — the same statement he made on October 12. His previously stated criterion deserves separate note: when robots can complete roughly 80% of tasks via voice or text instructions in 80% of unfamiliar scenarios — a rare testable claim in the industry; whether the moment has arrived can be measured, not argued. On October 25, China Daily published a follow-up: "Unitree ramps up push for AI breakthroughs."

8. Report on young embodied-intelligence scientists: mismatched centers of gravity

• Sources: People's Daily Zhejiang / Guangming Net (2026 World Young Scientist Summit)
• Link: http://zj.people.com.cn/n2/2026/1024/c186327-41731577.html
• Date: 2026-10-24

At the World Young Scientist Summit, Gao Chunbo (deputy secretary-general of the World Young Scientists Association) released the "Report on the Distribution and Mobility of Young Scientists in Embodied Intelligence," studying a sample of 1,836 researchers. Key numbers: the young research cohort averages 34 years old, with over 60% aged 35 or under; founders are relatively mature, with 57% aged 36–45; academic institutions account for over 60% of placements, with the top ten institutions concentrated in China and the US. Two conclusions juxtaposed reveal a mismatch: the US has the most young embodied-intelligence scientists, while China leads in the number of embodied-intelligence companies — talent density in academia, industrial density in companies; the two centers of gravity don't overlap.

9. Paper watch: tactile triple, world-model quadruple

No new arXiv batch on Saturday, so clusters from the Friday batch (cs.RO, 69 papers):

• Tactile hardware triple: OpenTac (open-source 3D tactile sensing ecosystem, customizable/extendable/transferable, 2610.31661), GelSlim 5.0 (single fingertip camera unifying vision and touch, 2610.31334), VecTile (low-cost flexible tactile array with 3-DoF force sensing, 2610.31472). Three tactile papers in one batch corroborates the competition finding that joint modules eat nearly half the cost — the first half of dexterous manipulation was models; the second half is skin.
• World-model quadruple: DiffDreamer (world model + differentiable simulation for agile quadrotors, 2610.31617), UniFlowWAM (unified world-action model with visual semantics and 4D geometry, 2610.30991), Skeleton-Grounded Latent Action Model (2610.31474), In-Context Learning World Models (bimanual deformable-clothing manipulation under partial observability, 2610.30988). In the same week DeepRobotics shipped DR-WM-1.0 — world models went from paper buzzword to standard industrial module in one or two quarters.
• Also worth noting: LocoManiLab, a unified testbed for humanoid mobile manipulation integrating data, policies, controllers, and deployment (2610.30876); and FocusVLA, which dynamically allocates camera viewpoints using future intent (2610.31423), a follow-up to yesterday's VLA coverage.

10. One-line briefs

• Human-robot fable play "Uninstall, Love" premiered in Shanghai — first stage play headlined by a humanoid robot (Sohu, 10-24)
• "2026 Embodied Intelligence Technology and Industry Development Report" released; one-line conclusion: "embodied intelligence's real industry inflection point has arrived" (EETimes China, 10-24)
• Zhiyuan released an embodied reasoning dataset — 150 hours of data teaching robots to do things (EETimes China, 10-23)
• Five stale stories blocked from recirculation today: Unitree IPO lottery win rate 0.0181% (August), robot half-marathon record (April), US ban on Chinese humanoid robot imports (late-July FCC case), Musk's 1 billion robots in a decade (September), Chengdu emotional-robot funding (10-16, already logged) — these headlines flooded timelines today; note the original publication dates

Editorial Observations

Money is starting to demand evidence. Today's three capital stories — raised IPO thresholds, data collection under scrutiny, orders shedding water — are three faces of one thing: for two years, embodied intelligence valuations rested on "story plus related-party transaction revenue"; now regulators ask whether revenue is real, markets ask whether orders hold water, and media ask whether data is genuine. Against yesterday's Unitree limit-up (624 yuan/share, market cap 252.39 billion yuan), the secondary market is still pricing stories while the primary market has already changed anchors — from imagination to cash collection. Regulators are running a stress test for the whole industry with a single question: strip away related-party transactions and subsidies — do your customers still pay?

While capital's patience thins, the technical chips thicken. DeepRobotics shipped five modules at once, a full stack from world model to data closed-loop; the same week brought an arXiv world-model quadruple and tactile triple; Zhiyuan sent 110 mass-production robots to dance in formation — testing exactly the fleet-scheduling capability that most resembles real work. Two curves are converging: the valuation line is searching downward for a foundation while the technology line builds floors upward. The industry report says "the real inflection point has arrived" — but an inflection point isn't declared; it's delivered by these two lines crossing. Until then, every day is an examination period.

Data infrastructure has become the shared answer on both sides of the Pacific. Musk's motivation for the "school" is blunt: autonomous driving has fleet data feedback; humanoid robots don't — so build 10,000–20,000 robots first and let them produce their own textbooks. This mirrors China's data-collection troubles: it's not that data is worthless, it's that fake data is crowding out real data's credibility. Whoever turns clean data supply into infrastructure holds the ticket to the ChatGPT moment Wang Xingxing describes. His criterion — 80% of tasks in 80% of unfamiliar scenes — is a rare falsifiable claim: whether it has arrived, just test it.