Embodied AI Daily · 2026-08-29

Today's Highlights

1. NDRC sets the deployment path for embodied AI: training grounds + pilot-testing bases as levers to tackle "data hunger," while warning against "blindly following trends and rushing in all at once" 2. Beijing's first humanoid robot data training center shuts down: Realman Robotics calls it a "data strategy upgrade" — centralized lab data collection 1.0 gives way to a real-world remote operation network 2.0 3. AGIC·GERX Shenzhen expo wraps up: 1,012 exhibiting companies, 1,200+ exhibits — "not a single concept model; all are priceable, orderable, and available for volume delivery" 4. Anatomy of 90 billion yuan in funding: 22 companies valued above 10 billion yuan; dexterous hands and data collection emerge as new favorites; Unitree's revenue structure fully disclosed for the first time (73.6% from research & education) 5. AI "super week" concludes: NVIDIA data center revenue hits $89 billion record; broker consensus on "compute + embodied AI" as the dual main investment lines

Detailed Content

1. NDRC: Solving embodied AI's "data hunger" while preventing blind expansion

• Source: NDRC August press conference (Beijing Daily / CFC / CLS multi-source)
• Link: https://xinwen.bjd.com.cn/content/s6a9155bde4b03fa51a8381d6.html
• Date: 2026-08-28
• Summary: Responding to questions after the World Robot Conference, NDRC policy research deputy director Li Chao laid out a roadmap centered on embodied AI training grounds and national AI application pilot-testing bases. Three supply-side elements: on data, "build high-quality real-robot data collection systems to solve the training data hunger"; on models, encourage exploration of multiple technical routes including VLA models and world models; on standards, "use unified standards to lower cross-embodiment model adaptation costs." Five priority application areas were named: manufacturing, healthcare, consumer services, services, and public safety. She also drew a clear line: the robotics industry must "develop in a healthy, orderly manner suited to local conditions, preventing blind trend-following and herd behavior."

2. Beijing's first humanoid robot data training center stops operating: data collection shifts from "centralized lab" to "real-scenario remote work"

• Source: Beijing Business Today (followed by Southern Metropolis Daily, Sina Finance, etc.)
• Link: http://www.bbtnews.com.cn/2026/0828/603992.shtml
• Date: 2026-08-28
• Summary: Beijing's first humanoid robot data training center — unveiled in March 2025 at Shougang Park, Shijingshan, spanning ~3,000 square meters with over 100 robots deployed — has ceased operations. Realman Robotics, the core equipment and technology provider, responded that the Beijing center was data collection 1.0: the centralized deployment completed zero-to-one validation and accumulated tens of millions of trajectory fragments, but "was still fundamentally a lab scenario." The 2.0 Changzhou Embodied AI Data Experiment Platform launched in May, built on the Global Labor Network (GLN) remote operation network, plans to deploy 100+ wheeled humanoid robots into real settings like factory lines and warehousing/logistics to complete over 1,000 real work tasks; 30+ robots are already deployed at scale. Robots collect data while "doing real work" — the company calls this a move from "effect demos" to "real-scenario, at-scale data delivery."

3. AGIC·GERX Shenzhen expo closes: "not a single concept model"

• Source: Shenzhen News / China Securities Journal
• Link: https://www.sznews.com/news/content/mb/2026-08/27/content_32157255.htm
• Date: 2026-08-26 to 08-28
• Summary: The 2026 AGIC Shenzhen International General AI Industry Expo and GERX Global Embodied Intelligent Robotics Expo was held at the Shenzhen World Exhibition & Convention Center: over 81,000 square meters, 1,012 exhibitors, 100+ global or China debuts, and 7,000+ overseas buyers from 30+ countries and regions. Xia Zuoquan, president of the Shenzhen AI Industry Association, said in his opening remarks: "Of the 1,200+ exhibits, not one is a concept model — all can be priced, ordered, and delivered in volume." SenseTime released the SenseNova U1 Pro multimodal foundation model during the event; iFlytek's Spark X1 and full-device lineup made their complete debut.

4. After 90 billion yuan in funding, where did embodied AI money go?

• Source: OFweek Robotics (Embodied Investment Research Circle)
• Link: https://robot.ofweek.com/2026-08/ART-898890-8420-30700428.html
• Date: 2026-08-27
• Summary: Domestic embodied AI funding in H1 2026 exceeded 90 billion yuan, already 67% of the full-year 2025 total. Valuation tiers: 22 companies now valued above 10 billion yuan, led by AgiBot at 36–45 billion; Zhi Ping Fang, Galbot, Self-Variables, Xinghaitu, and Qianxun Intelligent each stand above 20 billion. Since July, three dexterous-hand companies have reached 20 billion valuations (one barely a year old; the other two up 20x in 20 months). Unitree's IPO prospectus disclosed its full revenue structure for the first time: 2025 revenue of 1.699 billion yuan, net profit of 278 million yuan, and 5,215 humanoid robots sold — a global #1 — but research/education accounts for 73.6% of revenue vs. only 9.01% from industrial applications; revenue explicitly attributed to manufacturing/inspection/logistics was just 15.7 million yuan. A real-robot data pricing market has formed: 700+ yuan/hour, yet peak industry-wide output was only tens of thousands of hours in H1 — two orders of magnitude short of the tens of millions of hours general models need. The largest single order was several million yuan, with buyers mainly compute giants like Alibaba and ByteDance. Sequoia China made 12 investments across 16 leading companies in H1; "club deals" (commitments to lead the next round while investing in the current one) are among the mechanisms inflating valuations. AgiBot launched a Hong Kong IPO process in late July, Galbot completed its joint-stock reform, and Mech-Mind passed its HKEX hearing.

5. AI "super week" wraps: NVIDIA beats expectations, compute + embodied AI as dual main lines

• Source: China Securities Journal
• Link: http://intl.ce.cn/sjjj/qy/202608/t20260828_3174775.shtml
• Date: 2026-08-28
• Summary: NVIDIA FY2027 Q2 revenue was $96.22 billion (+106% YoY), with data center revenue of $89 billion (+117%, 92.5% of total), net profit of $59.69 billion (+126%), and next-quarter guidance of $108 billion. Three catalysts resonated this week: the AGIC expo, the Guiyang Big Data Expo (Aug 28–30, 372 exhibitors, featuring Huawei's Atlas 950 SuperPoD and SoftCom's Token Factory, with KuaWei Intelligence's second-gen general robot on debut), and NVIDIA's earnings. Broker consensus: AI is accelerating from cloud to edge and device sides, while humanoid robots move from prototype demos toward volume production. A-share compute-chain results are materializing: Foxconn Industrial Internet H1 revenue of 557.86 billion yuan (+54.6%), Zhongji Innolight net profit +241.7%, Inspur forecasting net profit growth of 226%–288%.

Editor's Observations

Observation 1: Data infrastructure "model iteration" is running ahead of "scale expansion." The most telling fact today is that two stories landed the same day: the NDRC ranked "building high-quality real-robot data collection systems" first in embodied AI factor supply, while the country's first landmark data infrastructure project — the Shougang Park training center — was reported shut down. Realman's answer: centralized lab collection is 1.0; "real work" data must come from real production scenarios. The Changzhou 2.0 GLN lets robots collect data as a byproduct of doing real work in factories — amortizing data collection cost into actual labor output, turning data from a "specially procured production material" into a "byproduct of labor." This matches the funding numbers: 700 yuan/hour real-robot data, tens of thousands of hours produced industry-wide in H1, versus tens of millions needed — a two-order-of-magnitude gap that can't be closed by building more collection sites, only by trading deployment scale for data scale. GLN is essentially teleoperation industrialized, at the starting point of the data production cost-collapse chain — but for the first time as "scaled commercial operation" rather than a "data project." The NDRC got the priority right; how training grounds should be built to avoid repeating 1.0's fate — Beijing's shutdown already gave that first lesson for free.

Observation 2: The industry's language has switched to "orders," but revenue still speaks "research." The expo's "priceable, orderable, mass-deliverable" framing, the NDRC's anti-herd warning, and Unitree's 73.6% research/education revenue share form a complete signal set: expos and regulators have both set "volume delivery" as the new baseline, yet real industrial penetration remains extremely early — Unitree's explicitly manufacturing/inspection/logistics revenue is only 15.7 million yuan against a first-day market cap that touched 444.9 billion yuan. The 90 billion yuan of primary-market capital is still searching for coordinates, and club-deal mechanics mean valuations still contain substantial institutional coordination; a starker contrast: Spring Festival Gala marketing spend (~100 million yuan each for four companies) vs. R&D (Unitree 145 million yuan for the year; several 20-billion-valued startups under 40 million). The single metric to watch next: the "industrial application revenue share" in AgiBot's Hong Kong prospectus — the hardest indicator of embodied AI crossing from research-budget markets into industrial markets. One easily missed detail: the NDRC wrote "lowering cross-embodiment model adaptation costs via unified standards" into an official statement — the model–embodiment adaptation layer is being planned as public infrastructure for the first time; interface standardization has entered the national agenda.

---

*Digest compiled by Su Mu; all sources verified against originals. Please credit the source when reposting.*