Embodied AI Daily Briefing · September 17, 2026

Today's Highlights

1. 'Revenue authenticity' becomes the industry's main narrative: Mech-Mind founder Shao Tianlan denounced 'assembled deal' revenue models and named Galbot; Galbot filed a police report in response; rumors of regulatory 'window guidance' surfaced the same week. 2. AgiBot GE-Act 2.0: real-robot data scaled from 300 to 30,000 hours, success rate 17.1% → 44.1% with no saturation seen — the first scaling curve for robot world-action models. 3. Skild AI released S1: a foundation model that learns new tasks from a single demo video, trained on NVIDIA's physics AI infrastructure. 4. Embodied AI IPO cohort expands in one week: Self Variables (Zi Bianliang) reportedly filed confidentially with the HKEX; Edge Medical (Jingfeng) started A-share listing tutoring; UDI Robotics rose as much as 176% on debut. 5. The 9th China Robot Summit held in Hangzhou; the 10th National Underwater Robot Competition kicked off in Dalian.

Details

1. Shao Tianlan vs. Galbot: The Industry Faces the 'Revenue Authenticity' Question

• Sources: Xinkuaibao (reporter Zheng Zhihui), 36Kr, iHeima
• Link: https://www.xkb.com.cn/articleDetail/524213
• Date: 2026-09-16

Timeline: Starting September 9, Shao Tianlan — founder of Mech-Mind, listed on HKEX less than ten days earlier — posted repeatedly on WeChat Moments accusing the industry of 'assembled deal' companies that fabricate 'fake, unsustainable revenue' through government-led 'data collection centers' and related-party transactions, lacking PMF yet seeking IPOs within two or three years — 'illegal, unethical, and not smart.' He named Galbot directly in the comments. On September 10, Galbot published a long post ('Embodied intelligence is a marathon; we only race against time') without direct retaliation; on September 11 an anonymous document alleging regulatory arbitrage in Galbot's HK listing circulated; on September 12 Galbot issued a stern statement calling the claims 'malicious rumors' and reported to police; on September 13 Shao clarified the report targeted the circulating PDF and was 'unrelated to me.'

Both sides are at capital milestones: Mech-Mind broke issue price on its September 1 debut, with cumulative 2023–2025 losses of RMB 1.044 billion (adjusted net loss RMB 657 million). Galbot has reportedly filed confidentially with HKEX, claiming thousands of cumulative industrial orders and 100+ space-cabin deployments, but has never disclosed full financials — the G1 priced at RMB 700,000, S1 around RMB 1.05 million, cumulative deliveries in the hundreds — a stark mismatch with its ~RMB 20 billion valuation, and it has formed joint ventures with local governments/state capital in Yibin, Hangzhou Qiantang, and Jingdezhen. Shao questioned the mechanism: 'Robots are sold to government-led data centers, then data is bought back through related-party transactions — the money circles among insiders and book revenue is created.'

Regulatory and market signals: On August 28 the NDRC's Li Chao warned against 'rushing headlong' into the sector; on September 9 market rumors said regulators had given informal 'window guidance' to investment banks: humanoid robot IPO candidates must prove sustained revenue, a loss-reduction path, or genuine technical innovation (unverified). In secondary markets: as of September 11, the Wind humanoid robot index traded at 28.37x PE (TTM); Unitree fell 55% cumulative over 17 trading days yet still traded around 350x PE, closing at RMB 469.80 on September 15 (–57% from its first-day high), while fair-value anchors from CITIC, CCB International, and Nomura (RMB 50.6–55.9bn / 109bn / 149.6bn) all sit below the current market cap. Haitong International's widely cited August 27 judgment: 'Demos will no longer be scarce; scaled manufacturing, data flywheels, and scenario capability form the long-term moat. Core competitiveness is shifting from the ceiling of product capability to the floor of commercialization capability.' Separately, on September 13 rumors circulated that Qianxun Intelligence, Galaxea, and Robot Era had been reported by peers for a 'key exam' this month; responses varied and all claims are unverified.

2. AgiBot GE-Act 2.0: 100x Real-Robot Data with No Saturation — A Scaling Curve for World-Action Models

• Source: AgiBot official release / Sina Finance
• Link: https://finance.sina.com.cn/stock/t/2026-09-09/doc-inirfcpy8693561.shtml
• Date: 2026-09-09

Unlike the mainstream 'off-the-shelf video generation model + action module' route, GE-Act 2.0's visual representation, future generation, and action prediction are all trained from random initialization on embodied manipulation data, with no fine-tuning on evaluation tasks — directly tested zero-shot on real robots in unseen settings (100 atomic tasks, 20 skill classes, two embodiments). A controlled comparison across 300/1,200/5,000/30,000 hours of data: scaling 100x lifted G1-OP overall success from 17.1% → 44.1% and non-zero tasks from 39 → 76, and G2-90D from 13.4% → 31.1%, with no visible saturation between 5,000 and 30,000 hours; fine manipulation like towel folding, cup nesting, and flower arranging only unlocked at the 30,000-hour tier. G2-90D, with under 2% of training data, gained 17.7 percentage points as shared data grew, validating cross-embodiment transfer. Architecturally: a control-oriented autoencoder (CoAE) compresses each 256×384 frame into 24 visual tokens (1/16 of DINOv3) with 97.95% instruction-frame matching accuracy; a one-step visual planner generates an action chunk in 104 ms on an RTX 5090; the KASO method resolves misalignment between future prediction and action supervision, adding 10 points of grasp success in unseen scenes. The fine-tuned version reached 60.52% success on RoboTwin unseen environments and 0.770 instruction following on GenieSim, surpassing π0.5 and GR00T N1.7. Total data: 39,000 hours of instruction-video pairs, 32,000 hours of robot trajectories, 30,000 hours of fully paired data. On September 8, Galbot peer Songyan Dynamics launched the Scalabot sub-brand and its action-conditioned HERON-World Model — a second world-model release in one week.

3. Skild AI S1: Turning 'Watch One Video' into In-Context Learning

• Sources: The Robot Report (Aug 31), NVIDIA blog (Sep 11), Skild website
• Link: https://www.skild.ai/blogs/learning-by-watching
• Date: Published 2026-08-31, continued buzz this week

Skild AI released its flagship robot foundation model S1, positioned as an 'in-context learner': given a single human demonstration video, the robot executes long-horizon tasks never seen before. Its thesis mirrors AgiBot's: Skild argues teleoperation cannot reach foundation-model scale (limited by diversity — data trapped in labs/deployments; limited by scale — real-time collection is too slow), whereas human video is the robot's 'Common Crawl'; the cost is missing signals (no torque/touch in video) and the embodiment gap. Key figure from the blog: new skills require video-watching fine-tuning plus ~1 hour of robot data. NVIDIA's blog confirms S1 was trained on NVIDIA's physical AI infrastructure. Skild's prior pitch: an 'omni-bodied brain' — one brain controlling any robot for any task — deployed in security inspection, mobile manipulation, and autonomous case picking.

4. IPO Cohort Expands; Two Payroll Ledgers

• Source: iHeima AI robotics weekly (2026-09-15)
• Link: https://timeline.sohu.com/news/q7EAMWiG0j
• Date range: 2026-09-07 ~ 09-11

Four capital moves: ① On September 10, Self Variables was reported to have confidentially filed with HKEX (company said 'unaware'; TMTPost had an exclusive on August 6). Founded December 2023 in Shenzhen with a full-stack end-to-end embodied foundation model (WALL-B), it has raised over RMB 5 billion across 10+ rounds (Meituan, Alibaba, ByteDance, Xiaomi, Sequoia, IDG, the national AI industry fund), completed B/B+/B++/C rounds April–June at a post-money valuation above RMB 20 billion; commercial landing at 58.com home services and a German luxury-brand auto parts line, with logistics sorting tested at 1,816 items/hour at over 98% accuracy. ② Edge Medical (multi-port + single-port + remote surgery robots, HK-listed in January) began A-share listing tutoring on September 11; 2025 revenue RMB 456 million (+184.8%), over 50% overseas. ③ UDI Robotics listed on the HKEX main board on September 9, briefly up 176% intraday; indoor/outdoor delivery and cleaning robots deployed in 20+ countries. ④ Estun completed a RMB 487 million all-cash acquisition of 100% of Kuzhuo (a related-party transaction, 901.20% appreciation), with a performance commitment of no less than RMB 620 million cumulative revenue May 2026–2029 — an industrial robot giant buying an embodied-AI startup.

Two payroll ledgers: AgiBot partner Yao Maoqing told Nanfang Daily: 1,500+ employees, ~RMB 1 billion in annual wages (~RMB 667k per capita), expanding to 2,000 by year-end; 2025 revenue RMB 1.05 billion, 75% on R&D, pursuing a HK listing. Unitree's 2025 annual report: 516 employees, average pay RMB 367k, revenue per capita RMB 3.2932 million, revenue RMB 1.699 billion, non-GAAP net profit RMB 591 million. Model companies run ~3x the headcount density and 1.8x the pay of hardware companies — two corporate forms competing on 'sustained revenue' on the same prospectus track.

5. 9th China Robot Summit in Hangzhou, Plus a 3nm 'Dual-Brain on One Chip' SoC Ecosystem Conference

• Sources: Chinese Mechanical Engineering Society, robotsci.com.cn
• Link: www.robotsci.com.cn
• Date: 2026-09-16 ~ 09-18

The 9th China Robot Summit is being held at Hangzhou's Future Sci-Tech City, focused on embodied intelligence, large models, and the robotics ecosystem. Hangzhou is no accident — it was the first Chinese city to legislate for embodied intelligence (the Hangzhou Regulation on Promoting Embodied Intelligent Robot Industry took effect May 1, 2026). At the same event, Guangzhou Huaruizhipu held a 'Dual-Brain on One Chip' 3nm SoC ecosystem conference: a single-chip fusion of AI perception/decision and hard-real-time motion control on one 3nm SoC, bundled with an upgraded physics-constrained world model and data-collection devices (vendor claims like 'globally leading' are unverified by third parties). 'Big-brain/small-brain fusion chips' were also among 44 new technologies showcased at last month's WRC — edge compute integration is moving from papers to product-launch cadence.

6. 10th National Underwater Robot Competition Opens Today: Real-Sea Venues with Live Underwater Broadcast

• Sources: Dalian Daily, DUT News, Science and Technology Daily
• Link: news.dlut.edu.cn
• Date: 2026-09-17 ~ 09-20

Guided by the National Natural Science Foundation of China and co-hosted by the Dalian municipal government and Pengcheng Laboratory, the 10th National Underwater Robot Competition opens today in Dalian, marking its tenth anniversary. September 17–18 features the ROV mission track at Lingshui Port; September 18–19 the ROV operations track and underwater communication track move to real waters near Xinghai Bay Bridge, with live real-time underwater broadcast of key seabed stages. Competition-as-data-collection, underwater edition: real-sea operations plus communication links are themselves test data.

7. Addendum: Leju Launches China's First 5G-A Dedicated SIM Cards for Embodied Robots (09-10)

• Source: iHeima weekly
• Date: 2026-09-10

Leju Robotics, with Anhui Telecom and Huawei, released the nation's first 5G-A embodied intelligence dedicated plan ('Jushen Yilian') and humanoid-robot SIM cards, deploying the first 5G-A park-wide humanoid group inspection demo in Hefei. The plan targets high uplink, low latency, and multi-robot concurrency: 500 Mbps peak uplink, ≤20 ms latency in 95% of scenarios, priority scheduling; priced at RMB 599 (standard) and RMB 999 (enhanced). For the first time, carriers treat robots as an independent billing category — the connectivity layer is starting to price embodied intelligence.

Editorial Observations

Observation one: Verification-bandwidth economics is spreading from the lab to the prospectus. Shao Tianlan's 'assembled deal revenue' accusation, translated: the weakest rung of the assertion ladder (book narratives built on related-party transactions) was sold to capital markets as the strongest (auditable financial fact). The 'sell to data centers, buy data back' loop mirrors academia's 'generation is free, verification is scarce' physics — the scarce resource is verification bandwidth, and this time investment-bank diligence teams and exchange reviewers are being asked to supply it. The most telling detail: even the 'cleanest sample,' Unitree, fails the strict gate — 73.6% of humanoid revenue in its prospectus comes from project-based, one-off research orders, inherently conflicting with 'sustained revenue' requirements; and a –55% drop in 17 trading days at 350x PE, with three broker valuation anchors below market price, shows the secondary market is already teaching the lesson itself. The recurring 2026 move, read alongside WRC disqualifications (control-log audits) and research-integrity pipelines: moving acceptance from the 'story layer' to the 'delivery layer.' The signal to watch next week: prospectuses from Galbot and Self Variables (if they reach hearings, related-party structures and revenue composition must be disclosed) — the industry's first systematic handover of auditable execution records.

Observation two: The same week, both ends of data economics placed their bets. AgiBot's GE-Act 2.0 bet on paired real-robot data: 300 → 30,000 hours, 17.1% → 44.1%, no saturation from 5,000 to 30,000 — the real-robot scaling curve hasn't peaked; every extra hour still buys performance. Skild's S1 bet on human video: video is the robot's Common Crawl, with new skills needing only video plus ~1 hour of robot data — trying to push data cost back to internet scale. Both flagships launched the same week, two sides of Xu Huazhe's ledger: 'One hour of data costs hundreds of yuan; a million hours is hundreds of millions — if it turns out this data was useless, it's a colossal societal waste.' GE-Act 2.0 is currently the hardest positive evidence for the 'useful' side, while it simultaneously prices Skild's route: with 5,000 → 30,000 hours still improving, it's not yet time to switch tracks. A hidden interface-tax story in the engineering details: CoAE compresses each frame to 24 visual tokens (1/16 of DINOv3) at 97.95% matching accuracy, generating an action chunk in 104 ms on an RTX 5090 — a control-oriented compressor preserving action-relevant structure and discarding the other 15/16, the same law as MKB and music tokenization ('place structure correctly and dividends emerge'), and the same deployment-collapse axis as edge FP8 engines. Zoom out: the world-model/action-foundation-model track has shipped again (HiDream/WALL-SS/Atlas → Songyan HERON → AgiBot GE-Act 2.0 → Skild S1) — the release density itself is another vote for the judgment that data and generation are solved, and the bottleneck is verification and deployment.