Embodied AI Daily Brief · 2026-10-31

Today's Highlights

• Zhiyuan Innovation Day: New metric "Worken" debuts, shifting robot business from "selling by unit" to "billing by work hours"
• Lingqu OS 1.0 released; 200 million lines of code to be open-sourced by year-end and donated to the OpenAtom Open Source Foundation
• Zhiyuan benchmarked GPT-6 Astra on 120 simulation tasks: near-zero success on force-touch tasks
• Unitree Q3 earnings analysis: revenue accelerating, non-GAAP profit contracting, R&D spending doubling
• Three coexisting H1 shipment figures: 40,000 units produced, 25,000 shipped globally, Zhiyuan self-reported 9,700

Details

1. Zhiyuan Innovation Day: From Selling Robots to Selling Worken

• Source: TMTPost (via Sohu Tech)
• Link: https://www.sohu.com/a/1094785603_116132
• Date: 2026-10-29 (Shanghai World Expo Center; reported 10-30)

Peng Zhihui delivered the scorecard: the 20,000th robot rolled off the line in September; H1 shipments were 9,700 units, global market share above 43%, and H1 revenue approached 2 billion yuan. But the talk's centerpiece was another line: "How many robots we produce does not equal how much productivity is actually added to the real world." Zhiyuan therefore built a new ruler — Worken (effective intelligent work hours), equal to active robot count × autonomous working hours × output efficiency ratio. The three multipliers answer three plain questions: how many units are truly on the job (not sitting in warehouses), how long they worked autonomously (not hand-guided by humans), and how many units equal one human worker. In the showcased scenarios, Chimelong's entertainment robots led with a Worken score of 33,480; Chimelong has deployed nearly 500 units, handled about 10,000 visitors daily during the National Day holiday, boosted Spaceship Park traffic by nearly 40%, and plans to expand to 1,000 units by year-end. The business model follows: partner Wang Chuang described customers "no longer renting robots by the day" but buying productivity packages billed by Worken consumed. The harder number is the human-to-robot ratio — Chimelong's first 300+ robots required hundreds of engineers, roughly 1:1; now it's down to 3:1, with targets of 10:1 by year-end and 20:1 next year. Wang's threshold is 20:1: only when a position adds just 5% to labor budget can the customer afford it.

2. Lingqu OS 1.0: 200 Million Lines of Code Open-Sourced, Donated to OpenAtom

• Source: TMTPost / Shanghai Securities News
• Link: https://www.sohu.com/a/1094785603_116132
• Date: 2026-10-29

Zhiyuan released Lingqu OS 1.0, the first embodied-AI-native operating system: plug-and-play support for over 100 peripheral types across 15 categories, and end-to-end latency jitter under 4 microseconds at 1000Hz high-frequency tasks. The open-source commitment is unusually strong: full open-sourcing by the end of 2026, contributing 200 million lines of production-validated code and over 5,000 ecosystem packages, donated to the OpenAtom Open Source Foundation to build a humanoid robot open-source community. Ecosystem stats: the Yuansheng program links 255 universities, 511 companies, and 80,000+ developers, with over 300 million yuan invested this year. Wang Chuang was blunt about the division of labor: Zhiyuan does only 0-to-1 scenarios, leaving 1-to-N to partners — "there's no need for us to earn every yuan."

3. GPT-6 Astra Benchmarked by Zhiyuan: Force-Touch Nearly Zero

• Source: TMTPost
• Link: https://www.sohu.com/a/1094785603_116132
• Date: 2026-10-29

On whether general-purpose large models will swallow embodied AI, Zhiyuan co-founder and co-president Yao Maoqing's answer: no, at least not now. In the cited testing, GPT-6 Astra achieved a 66.3% average success rate across 120 simulation tasks, with strengths in semantic understanding, action planning, and decision-making; but single inference averaged over 18 seconds, high-dynamics and fine manipulation were inaccurate, and force-touch feedback tasks had near-zero success — it never learned manipulation. His conclusion is a layered architecture: general models handle System 2 long-horizon planning; the middle layer holds skills trained via VLA and world models; the bottom is real-time motor control at 100–1000Hz. Co-president Xiong Yan added an economic figure: a minute-level inference call costs over a hundred dollars — "I don't believe anyone will deploy a 5-trillion-parameter model on a robot body." Yao's closing line is worth quoting verbatim: be on the AI side, not the side that gets pierced through.

4. Unitree Q3 Report: Three Numbers, Three Directions

• Source: Cailian Press / Market Insight / National Business Daily
• Link: Unitree Technology "2026 Third Quarter Report" (disclosed on the Shanghai Stock Exchange, evening of 10-28)
• Date: disclosed 10-28; analyses published 10-29 to 10-31

For the first three quarters: revenue 1.817 billion yuan, up 55.71% YoY; net profit attributable to shareholders 387 million yuan, up 270.54%; non-GAAP (deducted) net profit 349 million yuan, down 18.78%. Q3 alone: revenue 665 million yuan, up 69.93%; attributable net profit 113 million yuan, down 17.37% (down 49.55% QoQ); deducted profit 105 million yuan, down 17.42%. Three directions at once: revenue accelerating, headline profit soaring (mainly from non-recurring gains), profit contracting once those are stripped out — and multiple outlets noted slowing cash collections. R&D spending grew 249.58% YoY in the quarter and 191.48% for the first three quarters, mainly directed at model development — where the profit went, these numbers answer half the question. Market context: the stock has rebounded recently but has shed 163.4 billion yuan from its listing-day high. The market is buying the multi-year growth curve; the quarterly deducted profit gives the ground coordinates — the two lines don't currently coincide.

5. H1 Shipments: Three Ways to Count, Three Answers

• Source: MIIT State Council Information Office briefing (10-19) / IDC Global Humanoid Robot Tracker (9-29) / Zhiyuan Innovation Day (10-29)
• Link: https://www.sohu.com/a/1094785603_116132
• Dates: 2026-10-19 / 09-29 / 10-29

The same "H1 2026," three sets of numbers from three sources. MIIT (10-19): China's humanoid robot products exceeded 400 models, production surpassed 40,000 units, with the full year expected to top 100,000. IDC (9-29): global shipments neared 25,000 units, up 432.1% YoY, with Chinese makers holding over 95% share; market size exceeded US$740 million (up 322.7%), with roughly 19,000 units in China. Zhiyuan self-reported (10-29): 9,700 units shipped in H1, over 43% global share — back-solving from 43%, the denominator is about 22,600 units, which doesn't align with IDC's 25,000. In mid-August, another research firm reported 19,100 units (Zhiyuan 8,400, 44%; Unitree 5,900), 1,300 fewer than Zhiyuan's self-report. Production counts units rolling off the line, shipments count units dispatched, and self-reports count a company's own methodology — the number customers actually need, on-duty hours, is reported by no one. That's exactly the gap Worken aims to fill (item 1).

6. A-Share Humanoid Robots: Three Limit-Up Waves in One Week

• Source: Tencent News / THS closing data
• Link: https://view.inews.qq.com/a/20261026A0DBXR00
• Dates: 10-22 / 10-26 / 10-28

The robot sector hit collective limit-ups three times this week. On 10-22, eight stocks including Yifan Transmission, Saifutian, and Moons' Industries hit limit-up; on 10-26, Longte Smart hit the 20% limit within 3 minutes of open, followed by China Post Technology, with Huayang Intelligence and Fanya Microtex joining in the afternoon; on 10-28, over ten stocks including Ningbo Dongli, Wanxiang Qianchao, and Wuzhou Xinchun hit limit-up, the same evening Unitree released its Q3 report. The limit-up wave and the non-GAAP contraction unfolded the same week; the market is largely pricing expectations. A sober retail comment worth recording: go through the investor Q&A of limit-up stocks — few can show an actual complete robot product.

7. Paper Watch: Friday cs.RO Batch of 129 Papers, VLA Starts Counting Costs

From the Friday batch (Fri, 30 Oct 2026, 129 papers):

• Evaluation and cost: RoboSeed puts VLAs and world-action models in one multi-dimensional evaluation table, counting resource burden alongside capability (2610.40026); Quantize by Doing does task-aware mixed-precision quantization for VLAs (2610.38661); Decoding Geometry, Not Tokens has small VLMs directly output geometric trajectories, skipping token-by-token decoding (2610.39416). Read together: the VLA competition is shifting from "can it" to "can it afford to run."
• Humanoid whole-body: METATRON, a promptable whole-body execution layer (2610.38027); FORGE generates feasibility reference trajectories for agile-gait whole-body execution (2610.37974); Humanoid Dodgeball trains humanoids in fast dynamic obstacle avoidance via dodgeball (2610.40003); SONIC-K1 ports a general human motion tracker to small humanoids on a single consumer GPU (2610.39327); AcrossDEX 1.0 pretrains from human motion videos for cross-embodiment bimanual dexterous manipulation (2610.40210).
• World models: a continuous-action world model learns cross-embodiment spatiotemporal dynamics priors (2610.40401); PRIME W0 extracts physical priors from world models for manipulation (2610.39528); When More Planning Hurts pours cold water on "more planning is better" — learned planners can hurt world model performance (2610.37891); Seeing and Imagining as One merges video generation and 4D perception in one model (2610.39232).
• Data-collection hardware: UMI-Live, dual-arm data collection for in-the-wild scenes (2610.39385); Tactus, a smart glove capturing touch and hand pose simultaneously (2610.39076); TouchBridge transfers human demonstrations to robotic hands via tactile conditioning (2610.40349).

8. Quick Notes

• Peng Zhihui, same stage: physical AI is still early — the victory of physical AI is when 10,000 humanoid robots are on the job
• Old-news alert: a Cailian Press piece titled "Humanoid robot concept rally widens, Orbbec and 10+ stocks limit-up" is actually a July 31 story; today's searches for "humanoid robot limit-up" mix July and October items — check the dates
• Google ER 2 cloud access and Genie 4 native audio (noted in the 10-29 issue) have no updates; the overseas embodied line was quiet over the weekend

Editor's Note

This week's two big domestic events staged the same play: unit counts are up — then what? Zhiyuan's Innovation Day made the question explicit — customers no longer ask "can you build it" but "can I afford it" — and Zhiyuan's answer is changing the billing unit from units to Worken. Unitree's earnings confirm it from the other side: in the same quarter revenue grew 55.71%, deducted profit fell 18.78% — the faster the unit-count race runs, the more each unit's business needs proving. The three conflicting shipment figures are a symptom of distortion: production tells a capacity story, shipments tell a market-share story, self-reports tell a launch-event story — and the number customers want, on-duty hours, nobody reports. Worken fills exactly that gap. Harder than the formula is the human-to-robot ratio roadmap: from 1:1 down to 20:1. That 5% labor-budget threshold is the passing grade for this business moving from project-based to scaled — if it can't be crossed, even the prettiest metric is just launch-event language.

Industry and academia converged this week. Zhiyuan's GPT-6 Astra test — 66.3% success, near-zero force-touch, 18-second inference — gave the "will general models swallow embodied AI" debate its first concrete numeric boundary; the layered architecture (general models for planning, skills for execution, motors for real-time) is industry's answer. In the same Friday cs.RO batch, RoboSeed puts capability and resource burden in one evaluation table and Quantize by Doing quantizes VLAs task-awarely — academia is cutting the runtime cost of that middle skill layer. Both sides are asking the same question: can the customer afford this robot? The emergence of cost-aware evaluation dimensions and quantization papers says more about where the field is heading than another few points of success-rate improvement.

The limit-up wave and the deducted-profit decline played out the same week; both are real. The sector is pricing next year's story while the filings record this quarter's execution. That retail line — "few on the limit-up board can show an actual robot" — is the one most worth writing down from this rally.