Embodied AI Daily Briefing · October 15, 2026 (Issue #39)

Today's Highlights

• Tesla core supplier Xinjian Transmission disclosed in an IPO inquiry response that the Optimus design is near finalization and entering mass production, with weekly output up ~10x in recent months
• AishiDa's 100-store pilot results: 100 Zhiyuan Lingxi X2 robots served as in-store guides; store sales rose 39% YoY during the September campaign
• CIIF closing week's verdict: robots are no longer performing — they're actually working (Flexiv making liangpi noodles, ABB inspecting die castings, KUKA moving boxes)
• Hangzhou humanoids doing parts-filling work in factories; companies project two-to-three-year payback, just reaching the threshold for industrial bulk purchasing
• Consumer-level progress: Qiyuan T1 Pro hands-on review published alongside the question "are ten-thousand-yuan robots cyber toys?"
• UCLA PhoneBot uses an old smartphone as the brain; total bipedal robot cost under $500

Details

1. Tesla Optimus V3 design near finalization; weekly output up tenfold

Sources: 21st Century Business Herald / Jiemian News / Sina Tech

On October 12, Optimus core supplier Xinjian Transmission (Hangzhou) disclosed in its ChiNext IPO inquiry response that Tesla's humanoid robot design is nearly finalized and entering mass production. Optimus V3 is planned to start production before the end of 2026; Tesla has converted the Fremont factory's Model S/X lines into Optimus lines. Output ramp data: roughly a 10x increase in recent months, from dozens per week in Q2 2026 to hundreds per week now. Obstacles are stated plainly: supplier constraints, automation equipment failures, and the complex assembly requirements of the hand components make the year-end production target challenging. Xinjian Transmission is a national "little giant" specializing in planetary roller screws, achieving China's first mass production in 2022 — coinciding with Optimus Gen1's release. Screws are among the highest-value core hardware in humanoids. The weight of this news lies in its source: an IPO inquiry response is a disclosure document, not supply-chain hearsay.

2. AishiDa's 100-store pilot: sales up 39% YoY during the campaign

Sources: Jiqiren Dajiangtang / East Money

From September 28–30, cookware company AishiDa deployed over 100 Zhiyuan Lingxi X2 robots (store name "AiBao") across 100 stores nationwide. Robots worked at least 10 hours daily, roamed the whole store rather than holding fixed positions, with open livestreaming throughout. Numbers disclosed this week: store sales rose 39% YoY during the September campaign. Important context: the campaign also included three-city livestream events by Zhiyuan and AishiDa, so robot-guided foot traffic and livestream-driven traffic are mixed in the figure. This is the first discussable sales number from a retail scenario — robot sales associates have moved from "can they?" to "are they worth it?" Related: AishiDa's Intelligent Valley Embodied AI Innovation Center was unveiled in Taizhou on October 14, announcing an "R&D + mass production" platform for scaled deployment.

3. CIIF closing week: robots stop performing, start really working

Sources: 21st Century Business Herald / Guancha.cn / China News Economy

The 26th China International Industry Fair ran October 12–16 in Shanghai, with 310,000 m² of exhibits from 24 countries and regions. The closing-week assessment: humanoids and embodied intelligence are accelerating into real production scenarios, with competition shifting from motion demonstrations to solving real manufacturing efficiency and labor problems. On-site: Flexiv's robot making liangpi, ABB inspecting die castings, KUKA moving boxes, alongside Unitree, Inovance, AgileX/SLAMTEC peers (SL-Motor, AUBO ecosystem exhibitors), Dobot. Guancha.cn asked on October 13 what AI-in-the-factory still lacks — the answers pointed to data accumulation, maintenance systems, and scenario adaptation. Last year's expo keyword was "hardware"; this year's is "production lines."

4. Hangzhou humanoids "working" in factories: two-to-three-year payback

Sources: China.com.cn / China News Economy

On October 13, in a Hangzhou factory workshop, humanoid robots precisely filled empty part slots on a production line using robotic hands, working in coordination with the line. The economics: companies estimate two-to-three-year payback. That figure matters against industry benchmarks — per Cailian Press, industrial customers generally only commit to bulk purchases when robots pay back within 2–3 years. This pilot just meets the line, marking the first time one condition for bulk purchasing has been met. A recurring phrase in same-day coverage: "from dancing to working."

5. Consumer-level progress bar: Qiyuan T1 Pro review and the "cyber toy" question

Sources: Tencent Tech / PE Daily

Tencent Tech published a hands-on review of the Qiyuan T1 Pro (released September 20, priced 29,999 yuan, shipping from October 1; the Q1 standard version costs 19,999 yuan). Conclusion: "global embodied intelligence commercialization is currently in a transition from lab prototypes to affordable consumer products," with many vendors packaging actuation, perception, and large-model capabilities into small humanoids. The same day, PE Daily asked: are ten-thousand-yuan-level humanoids "cyber toys"? Read together, the first wave of consumer humanoid buyers receive not labor but an open question — vendors are packaging "capabilities," users are weighing "uses."

6. UCLA PhoneBot: old phone as brain, bipedal robot under $500

Sources: The Paper / Leikeji

UCLA unveiled the PhoneBot bipedal robot: 483 mm tall, 1.8 kg, 13 degrees of freedom, using a $23.90 servo per joint. Bill of materials excluding the phone is about $400; pairing with a used smartphone keeps total cost under $500 (~3,000 yuan). The core idea: the smartphone serves as the robot's compute and perception brain — an old phone is no longer e-waste but "a retired-at-factory industrial PC." Honor's Robot Phone (announced in August) follows the same vendor-side logic: a phone AI core controlling a robot chassis. Phone makers and robotics companies are converging on the question "where does the brain live?"

Paper Corner (new arXiv cs.RO batch)

• What Matters for World Modeling in VLAs (arxiv.org/abs/2610.19825): integrating world models into VLAs is popular, but methods differ simultaneously in world representation, integration, and policy backbone, making it unclear what actually works. This paper decouples and compares four representative paradigms — the field being hot enough to need a "take it apart" study is itself a signal.
• ACRO: Actor-Critic Rollout Orchestration (arxiv.org/abs/2610.19677): "when to continue vs. intervene" during VLA execution is usually a fixed rule; this paper makes it a learned decision — a Critic estimates success prospects from policy rollouts and selects intervention points meeting three conditions: more promising continuation, physical feasibility, and enough time to finish the task.
• Perturbation-Informed Execution (arxiv.org/abs/2610.19656): action-chunk VLAs predict a sequence of actions per step, and fixed execution prefixes waste later predictions; executing longer amortizes inference cost, but uniformly lengthening misses correction opportunities. This paper uses perturbation signals to dynamically decide execution length per segment — an engineering answer to the speed-vs-correction tradeoff.
• One-shot Human Video-Guided Adaptation for VLAs (arxiv.org/abs/2610.19746): adapting VLAs to new tasks typically needs task-specific robot demonstrations plus training — expensive and slow. This paper adapts from a single human video demonstration — human demos are an order of magnitude cheaper than robot demos, moving the data-source story one notch forward.

Editor's Note

Taken together, today's news is a batch of "money language" arriving on schedule. Optimus weekly output went from dozens to hundreds, a tenfold jump; AishiDa's store sales rose 39% YoY; a Hangzhou factory gave a two-to-three-year payback; PhoneBot pushed the whole robot under $500. Four numbers from four entirely different source types: an IPO disclosure, a retail campaign, a factory floor, a university lab. Last week's briefing said the industry had "two coexisting prices" — primary markets pricing by narrative, rent and secondary markets pricing by cash flow. This week's ground news collectively leans toward cash flow; the industry narrative is switching from "what moves can it do" to "how much, how fast to pay back."

But every number carries a footnote. First, hands remain the hardest bone: even as Optimus output rose tenfold, hand assembly was named a bottleneck — Tesla hasn't skipped the dexterous-hand hurdle, BYD is filing patents on dexterous hands, and the finest demo at CIIF was still making noodles. Second, the 39% happened during a livestream-supported campaign; how much of the robot guide's value is labor versus traffic event must wait for a month without livestreams. Third, the two-to-three-year payback just meets the industrial bulk-purchase threshold — meeting the bar, not exceeding it. Numbers appearing is good news; footnoting every number is clear-headed. Two concrete readings to watch next week: whether AishiDa's sales still grow in a no-livestream month, and whether Optimus weekly output stalls at three digits or climbs into four before year-end.

(Issue #39 · Daily embodied AI industry briefing · Collection window 10-14 09:00 to 10-15 09:50)