Embodied AI Daily Digest · September 26, 2026

Today's Highlights

• Regulators slow humanoid robot IPOs via window guidance: Unitree lost approximately 247.5 billion yuan in cumulative market value one month after listing; the review focus is whether government-related revenue reflects real demand.
• Mass production milestone data released: Zhiyuan (Agibot) shipped roughly 9,700 units to top the global ranking, with Unitree exceeding 7,000 in second place; Goldman Sachs' September report maintains a forecast of 890,000 units by 2030 — primary and secondary markets are pricing the same reality in two different ways.
• 2026 Yuelu Conference: The 10-billion-yuan Jinfurong Dachen Future Industries Fund landed in Hunan's Xiangjiang New Area.
• Metric-Bench (arXiv 2609.25841): An indoor spatial metric reasoning benchmark for VLMs; specialized training with verifiable numeric rewards improved general capabilities rather than degrading them.
• People's Daily front page on Hangzhou's 15th Five-Year Plan: Embodied intelligence industry chains based on manufacturing and logistics scenarios are clustering into industrial groups.

Detailed Coverage

1. Window Guidance Hits the Brakes: Humanoid Robot IPOs Enter the Audit Era

• Sources: Sina Finance / Reuters / Sohu Tech
• Links: https://finance.sina.com.cn, https://2aibot.com
• Date: September 21–24, 2026

Reuters reported on September 21, citing informed sources, that Chinese regulators are informally slowing the listing pace of humanoid robot companies through window guidance. Two review priorities: high valuations, and whether revenue from local-government-supported projects reflects sustainable commercial demand — if data-collection center revenue is excluded, some companies' growth stories need rewriting. On September 24, Sina Finance published a roundup titled "Unitree Loses 247.5 Billion; Humanoid Robot IPOs Hit the Brakes": Unitree Technology marked one month since its September 18 listing, having set a STAR Market record with a 0.018% lottery win rate; its cumulative market value loss now stands at roughly 247.5 billion yuan (per the report). Others awaiting scrutiny include Deep Robotics (racing toward the STAR Market to raise 2.503 billion yuan, reported 9/20), Galaxy General (per Bloomberg, has secretly filed with selected sponsors seeking up to a $4 billion valuation), and Zhiyuan (which has started a Hong Kong listing process).

Background context: criticism of "event-driven entrepreneurship" erupted around September 13 (a 36Kr review noted UBTech's Walker S won over 750 million yuan in data-collection orders, while Zhiyuan formed at least 30 joint ventures with local state capital, some with reverse-purchasing cases), media observed on September 16 that "Zhiyuan and peers began adjusting their narrative logic," and on September 20 the Mech-Mind founder publicly attacked event-driven ventures. Regulatory window guidance is the final link in this chain.

2. The Mass Production Divide: 9,700 Units vs. 890,000 — Two Readings of One Industry

• Sources: ifeng Tech / Goldman Sachs Global Physical AI Report (published September 1)
• Links: https://h5.ifeng.com, https://stock.caijing.com.cn
• Dates: September 16 / September 22, 2026

ifeng's September 16 roundup of shipment figures: Zhiyuan Robotics leads globally with roughly 9,700 units shipped, Unitree follows with over 7,000, and Galaxy General, UBTech, and Leju rank third through fifth. (Note: the headline claimed "two giants crossing 10,000 units" but the body figures are 9,700 and 7,000+ — go with the body.) Meanwhile, Goldman Sachs' September 1 global physical AI report saw a wave of coverage this week (Caijing, Shenzhen government site, September 22): global humanoid shipments are projected to grow from ~75,000 units in 2026 to ~890,000 by 2030 (up from a previous forecast of 256,000), and ~6.5 million by 2035. Deutsche Bank simultaneously raised its 2026 estimate from 17,500 to ~50,000 units and 2030 to 700,000. The report names logistics and warehousing as the first at-scale scenario, citing Amazon and JD as representatives — JD Logistics' five-year plan announced at the JDD conference in early September to procure 3 million robots (including 1 million autonomous vehicles and 100,000 drones) is the latest supporting datapoint.

There is an order-of-magnitude gap between investment bank projections and actual shipments. This gap doesn't mean anyone is wrong — the 75,000 figure is a global full-year forecast for 2026, while 9,700 units is a leading Chinese vendor's actual shipments — but it explains why, in the same week, analysts raised forecasts, regulators hit the brakes, and local governments distributed money: three actors each holding maps at different scales.

3. Yuelu Conference: A 10 Billion Yuan Fund Lands in Xiangjiang, Central China's "AI Into Factories" Template

• Sources: China News Hunan / Zhidx / The Paper
• Links: https://www.hn.chinanews.com.cn, https://m.thepaper.cn
• Dates: September 24–25, 2026

The 2026 Yuelu Conference was held in Changsha on September 23–24. At the opening ceremony, Hunan's Jinfurong Dachen Future Industries Fund was officially launched in the Xiangjiang New Area: 10 billion yuan total with a 3 billion yuan first tranche, dual-GP management, focused on AI, life engineering, quantum technology, and advanced materials, with embodied intelligence listed as an extended track. Zhidx's headline observation: "AI is moving from the screen into the factory" — humanoid robots, industrial internet, and embodied AI deployments were on display throughout the venue, as Changsha offers a central-China answer distinct from coastal cities: compete on scenarios, not parameters.

4. Hangzhou's 15th Five-Year Plan: Embodied AI "Industrial Clustering" Makes the Party Paper

• Source: People's Daily
• Link: http://cpc.people.com.cn/n1/2026/0926/c64387-40805783.html
• Date: September 26, 2026

People's Daily published "Hangzhou Accelerates the Channel from AI Research to Productivity," covering Hangzhou's 15th Five-Year Plan opening moves: embodied intelligence application industry chains based on manufacturing and logistics scenarios are clustering into industrial groups, with efforts to stimulate innovation among all actors while promoting safe and orderly industry development. Notable phrasing: "safe and orderly development" — in sync with this week's regulatory window guidance.

5. Metric-Bench: Filling the Spatial Metric Reasoning Gap With Verifiable Rewards — General Capabilities Improve Instead of Degrading

• Source: arXiv 2609.25841
• Link: https://arxiv.org/abs/2609.25841
• Date: September 22, 2026

Metric reasoning (how wide is this door, how far is that table) is a longstanding VLM weakness and a prerequisite for embodied tasks. Existing approaches use pixel-level supervision for spatial fine-tuning, often damaging general capabilities. This work takes a different route: Metric-Bench uses reference objects with known physical dimensions within images to guide models in implicitly learning 2D-to-3D mappings, without requiring camera intrinsics; the companion MetricReasoner applies structured prompting plus verifiable numeric rewards for reinforcement fine-tuning. Results: metric reasoning exceeds larger proprietary models by 43.1%; downstream embodied benchmarks improve by RoboSpatial +30.4% and ERQA +9.3%, while general benchmarks improve rather than degrade (V*Bench +15.9%, BLINK 88.9%). "Specialized training must hurt general ability" gets a counterexample here — the difference may lie in whether the reward is verifiable.

6. Briefs

• RoboTok (arXiv 2609.03199, September 2): An internet-scale data engine that, given a query video, retrieves manipulation-relevant human demonstrations from massive web video, centered on a latent motion space learned from 3D hand trajectories — robust across viewpoints, scenes, and occlusion. Hand-trajectory-based retrieval turns web video into a continuous supervision source for robot learning — adding another rung to the "zero-cost collection" end of the data spectrum: no collection needed at all, just index the internet.
• Benmo Power (06731.HK) lists September 29: The joint module layer rings the bell ahead of whole-robot companies; IPO price 21.6 HKD, market cap 7.98 billion HKD. The first IPO observation sample under the window guidance context — worth revisiting.

Editorial Observations

This week completed a closed loop: "claim standards" became "a pricing variable."

Nine days ago the chain went: Shao Tianlan questioned "unauditable revenue claims" → event-driven entrepreneurship criticism erupted → Zhiyuan and peers adjusted narrative logic → Mech-Mind's founder attacked event-driven ventures → regulatory window guidance stepped in. The key phrase of the final step: "what remains after excluding data-collection center revenue." Regulators are asking the same question this column keeps asking: who ran this number, under what methodology, is it reproducible. The claim-strength ladder reached its eighth rung: the capital market regulator.

The market is already answering. Unitree losing 247.5 billion in one month isn't purely sentiment: when a company's government-related revenue share can't be externally audited, the valuation discount is the market's "verification tax." Under this publication's verification-bandwidth economics, this line has long been on the ledger — generation is cheap, verification is expensive, and the bottleneck is always bandwidth. This week's new development: when verification bandwidth runs short, the secondary market invents a substitute (discounts), the primary market hits the brakes (window guidance), and regulators write the question into review checklists. Three mechanisms appearing in one week says more about a watershed than any research report.

The other side moved the same week: Yuelu launched a 10 billion yuan fund, and the party paper's front page featured "industrial clustering." Money hasn't left the industry — it's changing administrative levels, flowing from whole-robot valuation narratives toward fund pools and the infrastructure layer (Benmo's bell-ringing tomorrow being the latest datapoint). The center audits the numbers, localities add ammunition — this division of labor will persist for some time.

On the technical side, this week's highlight, Metric-Bench, offers a comparison: specialized training with verifiable numeric rewards improves general capabilities rather than degrading them. Specialization and generality need not be zero-sum — provided the reward is verifiable. A robot company's revenue structure works the same way: the more verifiable components (repeat purchases, third-party customers, deployment-time operating data), the harder the valuation. That's why regulators' first question is which revenue remains after exclusions.

Tomorrow's follow-ups: Benmo Power's first trading day, and any change in Deep Robotics' IPO review status.