Embodied AI Daily Digest - October 9, 2026
Key points
1. CCTV sets the tone: China's embodied AI industrialization is accelerating - 70+ training grounds operational, 2025 AI core industry scale exceeding 1.2 trillion yuan. 2. The Paper dissects 117 data collection projects: 67 confirmed doing real data collection, 27 still on the drawing board - the first systematic look at the "data collection business." 3. CAICT quantifies the data gap: globally available high-quality real-robot data is only in the hundreds-of-thousands to millions of hours range; a "ChatGPT moment" for robots requires tens of millions of hours. 4. Boston Dynamics CEO change: former Amazon Alexa/AGI head Rohit Prasad becomes CEO, effective October 7, 2026. 5. September humanoid funding: 49 rounds totaling ~10.8 billion yuan; the two largest deals were not for complete robots.
Details
1. CCTV and Economic Information Daily: industrialization accelerates, data is the "core fuel"
- Source: CCTV News / Economic Information Daily
- Link: https://news.cctv.com/2026/10/08/ARTIF7lg4qnNHkuoFMA7qaWM261008.shtml
- Source: The Paper · Alignment Lab
- Link: https://www.thepaper.cn/newsDetail_forward_34170041
- Link: https://www.robotics247.com/article/boston-dynamics-names-rohit-prasad-as-new-ceo
- Source: EET-China
- Link: https://www.eet-china.com/mp/a528868.html
- Link: https://finance.sina.com.cn/tech/digi/2026-10-05/doc-iniucpwm3098641.shtml
- Source: arXiv 2609.34010
- Link: https://arxiv.org/abs/2609.34010
- Singapore hosted its first humanoid robot fighting championship on Oct 7; the "competition-as-data-collection" model is spreading overseas.
- YYForce opened a robot training, data, and experience center in Singapore on Oct 6, targeting hospitality scenarios (PR Newswire).
- IROS 2026 (opened Sept 28, Pittsburgh) dexterous-hand aftermath: DH-Robotics' ADH-5-13 five-finger hand (13 active DOFs, 20 kg payload); Sharpa launched the D01 tactile manipulation robot and W02 ultra-compact full-tactile hand; DGTROBOTICS showed a fully direct-drive 13-DOF hand. "13 DOF" is becoming the new baseline; STCN says the dexterous hand market may boom.
CCTV reported (Oct 8) that China's embodied AI is moving from lab to industry, with training ground infrastructure, data aggregation, standards-setting, and real-robot deployment advancing in parallel. Over 70 embodied AI training grounds are operational nationwide, with 40+ under construction or planned, forming clusters in the Yangtze River Delta, Beijing-Tianjin-Hebei, and Pearl River Delta regions. China's 2025 AI core industry scale exceeded 1.2 trillion yuan, with over 6,200 AI companies. Reports also re-highlighted a fully domestically produced electronic architecture embodied robot unveiled in Yichang, Hubei in late September.
The Economic Information Daily (Oct 9) visited the Beijing Humanoid Robot Innovation Center's data and training base: nearly 6,000 square meters where 40 robot configurations practice sorting packages, folding clothes, grasping medicine, and more in realistic scenes (home, retail, industrial assembly, chemistry labs, elderly care). The base has delivered nearly 30,000 hours of high-quality data and houses ~200 sqm of optical motion capture - the same tech behind Tiangong Ultra's 8.64-second 100m at the World Humanoid Robot Games. The RoboMIND dataset, open-sourced in December 2025, has surpassed 20 million downloads globally.
2. The Paper's "Meishuke": dissecting 117 data collection projects
After a recently listed embodied AI executive publicly questioned the "data collection business" in early September, The Paper reviewed 117 data collection, training, and related projects from government filings and bidding documents: as of September 26, only 67 confirmed actual data collection activity, 23 primarily confirmed training/testing functions, and 27 remained in planning or construction. Only 46 of 117 project names directly reference data collection or production; 48 are called "training grounds/bases/centers." Capacity disclosure is a weak point - just 5 centers disclose capacity in hours, with combined annualized/planned capacity of ~1.4 million hours, but design capacity is not actual output. One center claimed "over 10 million data items annually" in June, then "24,000 training data items per day" in July, with no explanation. Local state capital is deeply involved (e.g., Suzhou Wuzhong district's state platform runs a data collection center; Wuzhong Financial Holdings launched an 11-billion-yuan robot industry fund). CAICT's sober judgment: training grounds are transitioning from a "construction boom" to a "capability consolidation" phase, with scene homogenization, limited actual capacity, insufficient data circulation, and no sustainable business model yet.
3. CAICT's Embodied AI Training Ground Report (2026): the official data-gap figure
The report states global usable high-quality real-robot data is only in the hundreds-of-thousands to millions of hours range, while tens of millions of hours are needed for a robotics "ChatGPT moment" - a one-order-of-magnitude supply gap. Beijing Humanoid's data lead Shi Mingwei: "Ten thousand hours of high-quality data may matter more than a million hours of repetitive data." The base's long-term goal is million-hour-scale "high-value, high-availability, verifiable" data production.
4. Boston Dynamics: the man who built Alexa now builds Atlas
Boston Dynamics named Rohit Prasad as its third CEO ever, after founder Marc Raibert (1992-2020) and Robert Playter. Prasad spent 12 years at Amazon, most recently SVP and chief scientist for Alexa and AGI. The appointment, effective October 7, comes ~9 months after Playter's departure, as Atlas shifts fully to electric hardware and commercialization. A motion-control icon handing the wheel to a large-model AI executive is a signal bigger than the personnel news itself.
5. September funding: 49 rounds, 10.8 billion yuan - the two biggest were not OEMs
Top deals: DroidBot (Digua Robotics) raised a $400M Series C (Mirae Asset lead, Meituan strategic co-invest, announced Sept 17), and SemiDrive raised ~1 billion yuan across C+ and C++ rounds - both chip/tooling companies, not robot makers. On the OEM side, LimX Dynamics closed a Pre-IPO round near $200M at a 15-billion-yuan post-money valuation, having raised $400M over six months.
6. FieldAI: $700M round in progress, valuation up 4x in a year
Per Business Insider, California-based FieldAI signed a term sheet for $700M at a $10 billion post-money valuation (vs. ~$2B over a year ago). Founded 2023, FieldAI builds a "general robot brain" - an intelligence layer for multiple robot types, not hardware. It would join Physical Intelligence (~$11B) and Skild AI in the top tier of "brain layer" companies. Note: this is a signed term sheet in progress; some Chinese reports prematurely call it completed.
7. ZeroBot: one photo to simulation, grasping learned from scratch in 119 seconds
ZeroBot, under "zero human demonstrations, zero policy pre-training, zero known object models," reconstructs a full object mesh from a single-view image via image-to-3D generative models, then runs RL in massively parallel simulation, using generated geometry plus a learned value function to sample contact states and accelerate training. Real-robot evaluation across grasping, pushing, articulated-object interaction, and multi-stage manipulation achieved 87% success with ~119 seconds average training time. This Real2Sim approach compresses scene modeling to seconds, effectively giving RL an infinitely resettable playground.
Briefs
Editor's observations
Today's two major domestic reports naturally complement each other. The Economic Information Daily shows how the "data fuel" is refined: 6,000 sqm, 40 configurations, 30,000 hours delivered. The Paper shows how much of the "refinery" is still on paper: 67 collecting for real, 27 in planning, 5 centers totaling 1.4 million hours of "design capacity." Read together, the true shape of the supply gap emerges - CAICT defined demand (tens of millions of hours), Beijing Humanoid defined delivery (30,000 hours), and the middle stretch of "planned capacity" is unauditable. The data collection business has never been questioned over "whether data is needed" but "who verifies the data." Shi Mingwei's line - quality over volume - is today's most honest statement, but it exposes the next bottleneck: there is no industry standard and no third party for the "data acceptance layer." As training grounds move from construction boom to capability consolidation, what must consolidate is verification capability, not floor space.
Money and people point the same direction. September's largest two checks went to chips and tooling; FieldAI, a non-hardware maker, quadrupled its valuation in a year to rival Physical Intelligence; Boston Dynamics gave its CEO seat to the man who built Alexa. The hardware layer is deflating (OEM valuations under pressure, IPO pace slowing) while the intelligence layer inflates - "capital betting on layers" has expanded from a Chinese pattern to a global one. Prasad's appointment is the most symbolic: a robotics CEO's job description is being rewritten from "can build machines" to "can make machines think." In the same week, dexterous-hand makers competed over 13 DOFs. Bodies compete on specs, brains compete on tokens - the market is betting on both ends, but the bet sizes are beginning to diverge.