Embodied AI Daily · 2026-09-14

Today's Highlights

• UBTech's world-first 10,000-unit-capacity humanoid robot super factory starts production in Liuzhou: one Walker S2 rolls off the line every 10 minutes
• China's National Data Administration hosts an embodied-AI symposium — all nine participants are from the "data side"
• CIFTIS wraps up: the full embodied-AI industry chain on display; CCTV frames it as "from debuts to fine-grained deployment scenarios"
• SEED-UMI: humans and robots wear the *same* exoskeleton to collect data, 3.0x more efficient
• Harness Robotic OS: a paper literally titled "Harness," where the LLM is just a swappable component
• Kailong High-Tech (300912) deploys embodied robots in a Wuxi university mobile service hall

Details

1. UBTech's Liuzhou Super Smart Factory goes live: 10,000-unit capacity moves from slide deck to takt time

• Sources: National Business Daily / Sina Tech / China News Service Guangxi (multi-source cross-checked)
• Links: nbd.com.cn, tech.sina.com.cn (domain-level attribution; search API does not return full paths)
• Timeline: commissioning ceremony 2026-09-12; concentrated coverage 09-13
• Summary: The world's first benchmark factory for industrial humanoid robot manufacturing designed for 10,000-unit-scale capacity officially opened in the Northern Ecological New District of Liuzhou, Guangxi. Covering 14,000 square meters, the line's design takt yields one Walker S2 every 10 minutes, with planned annual capacity exceeding 10,000 units. It was built through deep strategic cooperation between UBTech and Siemens Digital Industries Software, spanning overall factory planning, product lifecycle management, and digital-twin simulation; product lines are the Walker S series plus the Cruzr series. Supporting data points: the Walker S2 stands 1.76 m tall and integrates fourth-generation industrial dexterous hands (sub-millimeter fine manipulation); the previous line produced its 1,000th unit in Liuzhou last December — meaning the new factory jumps existing thousand-unit throughput straight to the ten-thousand-unit scale; and a Daiwa report in early September noted UBTech management reaffirmed full-year revenue guidance of RMB 3.5 billion (H1 revenue over RMB 1 billion, up more than 100% YoY).

2. National Data Administration embodied-AI symposium: the official narrative places embodied AI in the "data elements" domain

• Source: National Data Administration official site (full text verified)
• Link: https://www.nda.gov.cn/sjj/jgsz/jld/llh/llhldhd/0913/20260913104730557195636_pc.html
• Timeline: held 09-10; readout published 09-13
• Summary: Chaired by Party Secretary and Director Liu Liehong, with Deputy Director Xia Bing in attendance. Nine participants — the Institute of Automation, CAS; Beijing Institute for General Artificial Intelligence; Guanglun Intelligence; JD Group; ShenDu JiZhi; the Shijingshan Humanoid Robot Data Training Center; Galaxy General; Invariant; and the Beijing Humanoid Robot Innovation Center — offered views on "data empowering embodied AI development." The readout, for the first time, explicitly characterized the industry as "exhibiting data-driven characteristics, with continuously growing demand for high-quality, diverse, large-scale data," and pledged to "advance embodied-AI data standards in due course, guide local data systems to carry out related work in an orderly manner, and actively support enterprises in increasing data investment."

3. CIFTIS wraps: embodied AI goes from "exhibits" to "industry chain"

• Source: CCTV.com
• Link: https://news.cctv.com/2026/09/13/ARTIf8Ptr15WYR1Uz3Rr6bnJ260913.shtml
• Date: 2026-09-13 (closing day)
• Summary: CCTV's closing report: "the full embodied-AI industry chain gathered at CIFTIS as humanoid robots unlock multi-scenario applications." Hall 3 (telecom/computer/IT services) hosted over 130 companies and institutions, with humanoid robots showcasing the whole chain from whole-body locomotion and low-level control to data collection. The prior day's CCTV headline made it explicit: "embodied-AI industry-chain debuts move toward fine-grained deployment scenarios" — from showmanship to scenario granularity.

4. SEED-UMI: building "interface alignment" into hardware — human and robot wear one exoskeleton

• Source: arXiv 2609.11753 (submitted 09-10)
• Link: https://arxiv.org/abs/2609.11753
• Summary: The bottleneck in dexterous-hand imitation learning is collecting and transferring contact-rich demonstrations. Traditional wearable exoskeletons record only on the human side and retarget to the robot via open-loop calibration in free space — degrading at contact. SEED-UMI (Tengbo Yu, Hangxin Liu, et al.) has human and robot wear the same exoskeleton: joint encoders become a physically shared measurement, and the wrist camera sees the same mechanism in both human collection and robot deployment — retargeting shifts from a lossy software interface to paired cross-embodiment supervision. Policies train directly on raw exoskeleton-centric wrist images, no segmentation or inpainting needed. Across five contact-rich tasks, data collection is 3.0x more efficient than exoskeleton teleoperation, with 70.0% average deployment success.

5. Harness Robotic OS: layered runtime, LLM as a plug-in, self-evolution with a safety gate

• Source: arXiv 2609.11225 (submitted 09-10)
• Link: https://arxiv.org/abs/2609.11225
• Summary: HROS is a unified embodied-agent runtime for quadruped inspection, deployed on real hardware as Argos (residential-community patrol). The system has four layers: robot runtime / embodied autonomy skills / cognitive agent runtime / interaction and operations plane. A shared context couples physical state with agent reasoning; three-tier layered memory (working / episodic / semantic) retains operational knowledge; and a "safety-gated self-evolution loop" turns execution traces into versioned candidate updates, explicitly forbidding unconstrained online modification. The real-robot stack: Vbot quadruped + Fast-LIO2 localization + Hobot-Stereo depth + PCT-Planner/EGO-Planner planning + Qwen3-VL inspection analysis orchestrated by OpenClaw — the LLM occupies exactly one slot. Measured: 100% waypoint arrival, outdoor localization error <10cm, obstacle-avoidance response <200ms, hazard detection 85–95%, 99% success on alerting and structured reporting.

6. Kailong High-Tech (300912): embodied robots enter a Wuxi campus — a living sample of the cross-sector integrator path

• Sources: People's Finance (Securities Times) / JRJ.com
• Links: stcn.com (domain-level attribution)
• Date: 2026-09-13
• Summary: Kaiqi Embodied Intelligence, a subsidiary of A-share engine after-treatment company Kailong High-Tech (founded 2025-11, registered capital RMB 10 million, Wuxi Huishan), in cooperation with Jiangsu Mobile, has deployed an embodied robot in the Mobile service hall of Wuxi Vocational Institute of Commerce, handling mobile services 24/7 at 3–4 minutes per transaction. In June the two sides unveiled the "Kaiqi Embodied Robot Field Engineer Academy" for industry-education integration. With the parent's core business in after-treatment equipment, "embodied intelligent robots" are already listed as an in-operation product — an "A-share cross-sector capital + regional operator scenario + joint-venture subsidiary" integrator deployment path is taking shape.

Editorial Observations

1. Capacity narrative and valuation narrative diverge in the same week. Unitree closed at RMB 477.12 on 9-11 (Eastmoney data), down over 56% from its first-day high, market cap below RMB 200 billion — the same week UBTech hung a "one unit every 10 minutes" takt board on its Liuzhou line. Two narratives each cashing in: the secondary market keeps repricing on revenue quality, while the industry side answers "can it scale" with capacity ramps. Global H1 humanoid shipments were roughly 19,100 units, ~97% from Chinese vendors (industry data cited by China Daily earlier this month) — demand concentrated in China is real, but converting 10,000-unit annual capacity into revenue requires order absorption. Daiwa's reaffirmed RMB 3.5 billion guidance and Q4 order cadence are a harder test than factory takt. The structural contrast to remember: the old line took ~9 months to reach a thousand units; the new line targets ten thousand in a year. A 10x capacity leap meets a demand structure dominated by "planned procurement" (State Grid's May plan: ~8,500 units this year, ~RMB 6.8 billion investment). Whether the time gap between capacity and orders gets filled — Q4 will tell.

2. The Data Administration's participant list is the official version of the "data-collection spectrum." Of the nine symposium participants: Guanglun Intelligence does simulation data; the Shijingshan Data Training Center just completed a 4D Gaussian light-field upgrade last week; Galaxy General and Invariant represent the "data-system" narrative. The Data Administration defining embodied AI as an industry "exhibiting data-driven characteristics" mirrors the collection spectrum this digest tracks (teleoperation → competitions → trial use → crowdsourced home collection → mocap augmentation) — same thing, different subject: startups collapse collection costs while the state standardizes data elements. "Advancing embodied-AI data standards in due course" amounts to a policy entry ticket for the data-infrastructure track. The same day, SEED-UMI on arXiv gives this axis a new rung: its essence is not "a better exoskeleton" but eliminating the lossy software interface of retargeting altogether — the interface tax is paid once, in hardware (shared exoskeleton), buying 3.0x data-collection efficiency. And HROS adds a blunt sample to the harness>LLM theme: the paper is titled "Harness," Qwen3-VL is a replaceable inspection-analysis part, and governance (safety gating, versioned updates, no unconstrained online modification) lives entirely in the runtime — the "execution trace → candidate update → gated release" loop is the industrial-middleware version of last week's OpenRSI recursive self-improvement and Palantir's write-side trio. What was in papers last week is running on community-patrol quadrupeds this week.