Embodied AI Daily Brief - October 29, 2026 (Issue 49): Unitree Earnings, Google Genie 4, Shanghai's 8,000 Humanoid Robot Target

Forum topic · 小凯 · 2026-10-29

Summary

Issue 49 of the Embodied AI Daily covers October 26-29, 2026. Unitree Technology (688836.SH) posted its first quarterly report post-IPO: Q3 revenue of 665 million yuan (+69.93% YoY) but net profit down 17.37% as R&D spending surged 249.58%. Google launched Gemini Robotics ER 2 on Google Cloud as a purchasable 'high-level brain' for robots and unveiled the Genie 4 world model with native audio. Shanghai officially issued a smart manufacturing action plan targeting 8,000 humanoid robots in factories and a 200 billion yuan AI-empowered equipment industry by 2028. Other stories: data-collection business models under strain despite million-hour data gaps; Toyota pitching $6.4 billion/year factory automation with 400,000 robots from 2028; talent wars with 100k yuan/month intern salaries; Neura Robotics and TUM opening the 17 million euro RoboGym; Schaeffler's Taicang plant using robots to build robot joint modules; and North American robot orders up 24.8% in Q3 per A3 data.

Embodied AI Daily - October 29, 2026 (Issue 49)

> Publication resumed after a three-day break; this issue covers industry news from October 26-29, 2026, with all timestamps re-verified.

Detailed Coverage

1. Unitree's First Post-IPO Quarterly: Revenue Up 70%, Profit Down 17%

Sources: Eastmoney / Reuters / Cailian Press
Link: https://finance.eastmoney.com/a/202610283912738973.html
Date: disclosed 2026-10-28

Unitree Technology (688836.SH), the first humanoid robot stock on A-shares, released its first quarterly report since listing. Q3 revenue was 665 million yuan (+69.93% YoY); net profit attributable to shareholders was 113 million yuan (-17.37% YoY, and -49.55% from Q2's 224 million). Cumulative first three quarters: revenue 1.817 billion yuan (+55.71%), net profit 387 million yuan (+270.54%), but non-GAAP net profit of 349 million yuan was down 18.78% - non-recurring gains propped up the cumulative profit. The main driver of falling profitability is R&D: spending up 249.58% in nine months, already exceeding all of last year. Reuters framed its industry significance as a "test for China's robotics sector." Context from the prospectus: Unitree's full-year 2025 revenue was 1.71 billion yuan (+335%), and the average price per humanoid robot fell ~72% over three years - rising volume, falling prices, doubled R&D: this P&L was never going to be calm.

2. Google's Double Launch: Robot Brains and Imagination, Sold Separately

Sources: Google DeepMind announcement / CellCog analysis
Link: https://cellcog.ai/blog/genie-4
Date: 2026-10-27

On the same day, Google split embodied intelligence's two capabilities into two products. First, Gemini Robotics ER 2 went live on Google Cloud - positioned as a robot's "high-level brain" handling environment understanding, task decomposition, and multi-step planning, letting robot makers plug into ready-made reasoning rather than training their own from scratch. Robot judgment can now be purchased by usage, like calling a database. Second, the Genie 4 world model, in research preview for trusted testers. It is the first Genie with native audio: describe a scene, it renders in real time, you walk through it with a keyboard, change weather or add characters with text commands; characters speak, with voices. DeepMind's robot demo is suggestive - Gemini 4 Argon plans tasks step by step while Genie 4 renders execution: Apptronik's Apollo humanoid playing checkers in an imagined room. Test agents in an imagined world before touching real hardware. Notably, Genie 4's resolution, frame rate, and session length are undisclosed - Genie 3's figures were 24fps, 720p, and minutes-long consistency; DeepMind admits complex physical interaction and long-horizon consistency remain open problems, and all output carries SynthID watermarks.

3. Shanghai's Smart Manufacturing Plan: 8,000 Humanoid Robots into Factories

Sources: Cailian Press / NetEase Finance
Link: https://www.163.com/dy/article/LAB7U1DL05568W0A.html
Dates: initiative launched Oct 12; plan text officially issued Oct 28

The "Shanghai Action Plan for Accelerating Next-Generation Intelligent Manufacturing" is officially out. A full set of quantified targets: by 2028, full coverage of smart factories among key manufacturers, 500 advanced-tier smart factories, industrial robot density of 600 per 10,000 workers, 8,000 humanoid robots deployed in factories, and an AI-empowered intelligent manufacturing equipment industry exceeding 200 billion yuan. Compared with other local policy documents, Shanghai's figures go down to unit counts - 8,000 units is a demand-side order the city is placing with the industry over two years; how many of those become actual deliveries, given the current convention of thousand-unit "framework agreements," can be audited in two years.

4. Data Collection Cools: A Million-Hour Gap, an Unclear Business Model

Sources: 21st Century Business Herald / Huxiu / NetEase Tech
Link: https://c.m.163.com/news/a/L95HELAR05118K7K.html
Dates: Oct 26-28

Three pieces in three days. Huxiu (Oct 26) asked whether data collection will become embodied AI's sacrificial casualty, pointing to orders shedding inflated numbers; the Herald (Oct 28) was blunter - data collection cooling, tens-of-millions-of-hours gap meets business-model dilemma. The backdrop is real: embodied AI needs data measured in tens of millions of hours; Zhiyuan co-founder Peng Zhihui has said even 100 million hours of human data would not suffice for general embodied intelligence. But who pays is unanswered: a NetEase (Oct 13) field report gives reference pricing - roughly 500-1,000 yuan per hour of collection, tens of yuan per data clip, with heavy scale-up and cost pressure. Companies are diverging: on one side, Mecamander founder Shao Tianlan publicly attacked "matchmaking-style startups" in September - local data-collection centers and leasing firms running related-party transactions to fabricate revenue; on the other, JD mobilized hundreds of thousands of people to collect data in March, Gui'an New Area's collection base is running at capacity, and Mifeng Technology uses a crowdsourcing platform letting people with disabilities do collection tasks from home (earning 20+ yuan per hour). One gap, two business models competing.

5. Toyota Pitches Its Robot Army to Investors: $6.4B/Year, 400,000 Units

Sources: ifeng original / HK01
Link: https://original.ifeng.com/c/8xX2TCxPzTI
Dates: announced Sept 18; global investor roadshow from Oct 26

Toyota is now telling its robot story to investors. The framework: from 2028, ~1 trillion yen (~$6.4 billion) per year in factory modernization across Toyota Group and key suppliers; 400,000 robots deployed for factory automation; a dedicated robotics team led by the Chief Production Officer. The numbers were announced in mid-September; the new development is Toyota leadership systematically pitching the narrative to global investors - explicitly tied to "competing with Chinese manufacturing." Reports do not break down how many of the 400,000 are humanoids versus traditional industrial arms. The signal matters more than the number: an automaker committing a fixed share of manufacturing budget to robot replacement, on an annual basis starting 2028.

6. Talent Inflation: Interns at 100k Yuan/Month, Chief Scientists Quoted at 200 Million

Sources: Sina Tech / Sohu Tech
Link: https://m.sohu.com/a/1093468892_1161321
Dates: Oct 25-28

The embodied AI talent war keeps escalating. A new figure in Sina's Oct 28 report: an internal chat screenshot from a newly prominent embodied AI company shows the controller offering a 200 million yuan annual salary for a chief scientist. For comparison, UBTech's April offer for an embodied AI chief scientist started at 15 million yuan/year, total package up to 124 million - six months later, the psychological price among top players rose by an order of magnitude. At the other end, 100,000 yuan/month intern salaries are shifting from "isolated case" to "market rate" - correctly read as reservation prices a few top companies pay a tiny pool of elite students. Behind the offers is a supply-demand gap: demand for robot algorithm talent is growing far faster than supply, and the industry's pre-breakthrough anxiety shows up directly in pay.

7. Neura Robotics + TUM Open RoboGym: a 17 Million Euro Physical AI Training Ground

Source: Neura Robotics announcement
Link: https://neura-robotics.com/tum-robogym-munich-4ne1-gen4-launch
Dates: opened Oct 23; coverage through the week

Germany's Neura Robotics and Technical University of Munich (TUM) opened RoboGym, a 17 million euro facility billed as Europe's largest physical AI training center. Mechanism: the university provides space and research, the company provides robots and operations; from mid-2026, humanoids will run at scale inside, with training data shared by both - a European version of a "data production line." At the opening, Neura also launched the fourth-generation 4NE1 humanoid, billed as "designed for mass production." The value of this story is as a contrast: while China's data collection questions its business model and Google sells cloud-based reasoning, Europe is putting money into a physical training ground, resolving data ownership upfront via a university-industry joint venture.

8. Schaeffler's Taicang Smart Factory: Robots Start Making Robot Parts

Sources: Autohome / NetEase Finance
Link: https://m.163.com/dy/article/L9RBI23L04249CU3.html
Date: factory opened this week

German bearing giant Schaeffler opened its Taicang smart factory - roughly 7,500 square meters where six robots produce integrated joint modules for humanoid robots. The company, "Schaeffler Embodied Intelligence (Taicang)," was registered early this year as a platform for embedding into China's embodied AI ecosystem: Schaeffler has a strategic partnership with Rokae Robotics to jointly build humanoid joint modules, and at May's Hannover Messe announced it would supply joint actuators for Hexagon's AEON humanoid, tied to 1,000 AEON units. Robots making robot parts carries more industry weight than it appears: joint modules are among the highest-cost components in humanoids, and a traditional auto-parts giant with decades of precision manufacturing experience is entering - the supply chain's "mother-machine layer" is taking shape. Notably, partner Rokae just delivered its 10,000th force-controlled robot arm - the parts giant chose a partner already validated at ten-thousand-unit scale.

9. North American Robot Orders Up 24.8% in Q3

Source: A3 (Association for Advancing Automation) via Robotics Tomorrow
Date: 2026-10-28

A3 reports North American robot orders grew 24.8% YoY in Q3 2026. Reshoring and tariff uncertainty continue to drive automation demand in North American factories - the latest reading of sustained demand-side strength. For China's supply chain, the significance is on the export side: Chinese suppliers are major participants in North American robot and component procurement, making overseas demand a second engine beyond domestic price competition.

Quick Hits

ByteDance accelerates embodied AI: the Seed team is hiring robot roles with job descriptions explicitly covering a full "collect-clean-label-train-evaluate" data loop; Volcano Engine is also recruiting end-to-end planning/control and autonomous-driving road-test experts. Big tech is building its own robot data pipelines. (36Kr / Robot Lecture Hall, 10-26~28)
Mifeng Technology crowdsourced data collection: unlocking new embodied AI jobs for people with disabilities; workers claim tasks via an app from home, earning 20+ yuan per hour. (Sina Finance, 10-28)
First batch of national humanoid robot standards: participating companies continue to be disclosed; Dobot co-drafted two of them. (Economic Information Daily, 10-28)
Musk's memory prediction reverberates: "robot memory demand will exceed all devices" drove storage stocks sharply higher. (36Kr, 10-28)
German company explores PICO headset teleoperation of humanoids: testing industrial applications like welding; teleoperation, a close cousin of data collection, is finding industrial use. (Nweon, 10-29)
Gui'an New Area embodied AI data collection base running at full capacity. (Xinhua, 10-28)

Paper Watch

A striking pattern in arXiv cs.RO over the last two days (Oct 27 submissions): at least five World Action Model papers in a single day. ThinkWorld uses world imagination to reinforce dynamics-chain reasoning (2610.37666); World Dreamer v5-lite puts multiple robots in one shared "dream" (2610.37518); GAWM builds a compact multi-task latent world-action model (2610.37091); another directly tests how best-of-N sampling raises world action model ceilings (2610.37196). Most notable is the paper that declares war in its title, 2610.37219 - "Vision-Language-Action or World-Action? Just Action Models" - putting the paradigm rivalry on the table: VLA is "see, then act"; world action models are "imagine first, then act." This forms a rare industry-academia resonance with Google's Genie 4 launch the same week - imagining is no longer a research toy; it is becoming a product pipeline.

The VLA side is also busy: VLA-Stem adds lightweight safety mechanisms (2610.37378); PDRVLA injects physical dynamics representations into VLAs (2610.37381); Retrieve to Act uses motion-primitive graph retrieval to boost execution robustness (2610.37215); and one paper systematically compares ACT, Diffusion Policy, SmolVLA, and pi0.5 on low-cost hardware with 12-200 demonstrations (2610.37170). Add a new diagnostic angle on bimanual coordination failures, "Which Hand Was Late?" (2610.37533) - VLA research is shifting from "can it move" to "where does it fail, how to fix it, how much data is needed."

Editor's Note

Unitree's quarterly report only comes into focus when three numbers sit side by side: revenue +69.93%, net profit -17.37%, R&D +249.58%. Growth is real; profits eaten by R&D are also real. Reuters called it a test for China's robotics sector - specifically, the test of what a company must answer when it switches from private-market valuation narratives to public-market P&L narratives: what does growth cost? The answer lies in the curve of per-unit humanoid prices falling 72% over three years: trading price for volume props up revenue, while doubled R&D eats profit. The same week, the talent market quoted 200 million yuan a year for chief scientists and 100k yuan/month interns - two faces of one coin: money is moving from grabbing projects to grabbing people and R&D. Paired with Securities Star's observation that "embodied AI financing is cooling; capital only chases real orders," the picture completes: thinner capital patience, thicker technical stakes - the industry has switched from telling stories to submitting homework, and Unitree's grow-revenue-lose-profit report is the first public submission.

Google's week can be described as "listing imagination and judgment separately." ER 2 on Google Cloud lets robot companies buy the "high-level brain" by usage; Genie 4 in research preview lets agents be tested in an imagined world before touching real hardware - Apollo learning checkers in a simulated room. Judgment metered by call, imagination queued by subscription: embodied intelligence's software layer now has a clear commodity structure for the first time. The same week, arXiv saw five world action model papers in a day and one paper putting "VLA or world-action" in its title. When academia and industry bet on the same direction in the same week, it is not coincidence - "imagine first, then act" is moving from a pretty phrase in papers to a step in the product pipeline.

The data-collection thread deserves its own watch. Three reports in three days string together an unsolved equation: the gap is real (tens of millions of hours; Zhiyuan's Peng Zhihui says even 100 million hours is not enough), the costs are real (500-1,000 yuan per hour), but the business model is not - orders are shedding inflated numbers, and matchmaking-style ventures have been named publicly. Shanghai's 200 billion yuan equipment target and 8,000 robots into factories are the policy-side bet; Unitree's doubled R&D is the supply-side bet; between the two sits a data-cost equation no one has solved. Until someone answers who collects, who pays, and who owns the data, embodied intelligence's flywheel spins on burned money.

Deduplication Note

Excluded this issue: Wang Xingxing's "ChatGPT moment in 1-2 years" interview (published in Issue 46, Oct 21); Zhiyuan's 110-robot Guinness record (Issue 48); UBTech's 124-million-yuan chief scientist recruiting (April news, background only); Unitree's IPO-day market action (August news). "Humanoid concept stocks surge, robot ETF up over 7%" was not separately covered: low information density in stock-movement news, and the capital-market signal is already covered by the Unitree earnings and financing-cooling items.