By Sarah Xiao 
The company registered on 11 February 2026. By June it was telling reporters it held nearly ¥1 billion in orders, had received ¥110 million in financing, and carried a ¥1 billion angel-round valuation. In September it launched an explosion-proof quadruped it described as the highest-rated in the industry. In October, according to Chinese business outlet Yingzi Beiwanglu, roughly ¥8 million in wages had gone unpaid, social insurance and housing-fund contributions had lapsed, the site had stopped, and the head of R&D was sitting in a labour-arbitration hearing next to ordinary staff.
Eight months, from model project to arbitration docket.
The uncomfortable part is not that a startup failed. It is that Wujing Dynamics (无境动力) executed, almost line by line, the standard script the embodied-AI sector has been running for two years — elite university team, hard industrial demand, state-owned-enterprise framework orders, local government fund investment, a differentiated explosion-proof heavy-load product — and then hit a trap buried inside every one of those lines.
The pitch was not the problem
On paper the configuration was close to a template. About twenty core members from Tsinghua, Shanghai Jiao Tong, Zhejiang and Harbin Institute of Technology, several with stints at Zhijiang Lab, AgiBot and Robot Era, covering joint modules, motion control and model training. The product choice was sharper than most: not a general-purpose humanoid, not a consumer toy, but “inspection plus manipulation” for industrial quadrupeds — a robot dog that carries an arm and does work while it walks, aimed at power grid, petrochemical and mining customers who pay six figures per unit and are often state-owned.
The September launch, the P25-Ex, claimed IIC T6 explosion-proof rating at 60 kg total mass, aimed at refineries and coal-chemical plants where a spark is a catastrophe.
Every element is defensible. The problem is that a defensible slide deck and a defensible business are different objects.
The arithmetic never closed
¥110 million sounds like a lot until you price the work it was supposed to buy: product freeze, explosion-proof certification, a production line, and delivery against a claimed ¥1 billion order book — simultaneously, within months.
The sector’s own balance sheets show what that costs. Sundry Robotics (云深处), one of the most commercially advanced quadruped companies, reported ¥337 million of revenue in 2025 and its first annual profit, at ¥28.68 million net. In the first half of 2026 revenue reached ¥259 million, more than double year on year, and the company was back to a ¥8.81 million net loss — with quadruped gross margin falling from 54.35 percent to 39.75 percent, and six humanoid units sold in six months.
Unitree sold more than 23,000 quadrupeds in 2025, with revenue roughly five times Sundry’s, on the strength of consumer and research volume that pushed unit cost towards ¥10,000.
That is the ledger the leaders produce with real cash and real shipments. Robot dogs can make money — slow money, hard money, scale money — and the margin can flip back to a loss when the mix shifts. A company with ¥110 million trying to complete product definition, certification, line build and delivery of a billion-yuan order book in one season has no buffer in any direction.
The word “order” was doing the work
The ¥1 billion order book split in two, according to the report: roughly ¥500 million in framework agreements with state-owned enterprises, and roughly ¥500 million in “local market reserve orders”.
Framework agreements have a specific meaning in this industry, and a filing from the same period makes it explicit. When Qiteng Robotics placed a ¥943 million framework contract for quadruped components with Meihu, the listed counterparty told regulators in writing that the figure was “the maximum possible order amount, estimated from the maximum purchase quantity in the contract”, and that it was “not binding as to the timing of order placement or the quantity of orders”.
A framework contract is an intention, a ceiling and a memo that leaves both sides room to walk away. Presenting it as revenue is a category error with a schedule attached: it converts a maybe into a hiring plan, a production line, a lease and a payroll.
| What gets called “an order” | What it legally is | What you can spend against it |
|---|---|---|
| Framework agreement | Ceiling, non-binding on timing or quantity | Nothing until a purchase order arrives |
| Letter of intent | Non-binding intention | Nothing |
| Pilot / evaluation unit | Small paid deployment, may not repeat | The invoice only |
| Purchase order | Binding commitment with delivery terms | Working capital, with delivery risk |
| Repeat annual order | Evidence of production demand | The thing investors actually want |
Capital moves in months. Sites move in years
The funding data explains why the timeline got compressed.
In the first half of 2026, Chinese embodied-AI companies closed 322 rounds worth ¥93.5 billion, five times the year before. Quadruped funding alone went from ¥300 million in 2023 to nearly ¥4 billion in 2025 and about ¥2.4 billion in the first five months of 2026, with the average round rising from ¥51 million to ¥299 million.
Demand did not accelerate at that rate. Most current orders are evaluation purchases — a few units to try — not expansion purchases driven by productivity. Between the two sit reliability validation, a service organisation and an ROI calculation that a customer’s finance department has to sign.
Sundry spent seven to eight years in grid inspection before more than a hundred substations under State Grid and China Southern became stable accounts. A company four months old announcing that state-owned framework orders had “entered the delivery phase” is describing a phase whose content — what was delivered, how many units, against what acceptance criteria — cannot be cross-checked in any public report.
The consumer side is cooler still. China sold about 37,700 consumer quadrupeds in 2025, up nearly 290 percent, but education and home use accounted for roughly 60 percent of units and less than a third of revenue. Prices have fallen below ¥10,000 — Vbot’s founding-member price was ¥9,988 with more than a thousand preorders on day one; another brand reached the top of a JD.com category in four months with over 2,000 consumer buyers. Cheap enough to buy is not the same as useful enough to keep using.
The incentive structure, not the villain
Local-government funds and industrial parks are not a defect in China’s robotics story. Unitree and Sundry grew up under versions of the same logic: district capital, a park, patient money.
What the Wujing case exposes is a tempo mismatch inside that logic. When the scorecard rewards “landing and exploding” in the same fiscal year, a company is pushed to compress three years of validation into six months of reporting. The fund gets a signed project, the park gets an occupied building, the company gets a valuation it cannot grow into — and the invoice for all of it arrives later, at payroll.
According to the report, Jiangning district invested several tens of millions of yuan through a district fund, the Huarui embodied-AI park invested a similar amount, and the company was given a standalone R&D building fitted out down to the desks. The district’s investment arm manages more than ¥12.3 billion across four funds, including a ¥1 billion science and technology fund with 46 portfolio companies. None of those numbers is unusual. That is exactly the point.
There is one more detail worth noticing: a fund whose investment was disclosed in June and partly paid in, while the shareholder register had already changed by 24 August, with no public clarity on whether the round closed or the terms were re-cut. Hesitation between due diligence and wire transfer usually arrives earlier than unpaid wages. It is just quieter.
What to check before the next quadruped term sheet
| Question | Why it decides the outcome |
|---|---|
| How much of the order book is purchase orders, not frameworks? | Framework value can be 100 percent of the headline and 0 percent of revenue |
| What is the burn multiple per delivered unit? | Delivery-heavy hardware eats cash before it earns it |
| Has the certification been granted, or applied for? | Explosion-proof certification gates the only high-margin segment |
| How many units are in continuous service, and for how long? | Pilot units and fleet units are different businesses |
| Does the founding team’s background match the product’s bottleneck? | Optics teams and robotics teams fail differently |
| Who signs off on service response time? | Industrial buyers purchase uptime, not capability |
The line worth keeping
The sector is not broken. Chinese companies build more than 90 percent of the world’s quadrupeds, industrial penetration has barely started, and falling prices are opening a consumer market that did not exist five years ago. None of that guarantees that every company on the track survives — and that has never been how tracks work.
What the last eight months suggest is narrower and more useful: the failure mode is rarely a missing technology. It is a language problem. “Order book” blurs framework and revenue. “Star team” hides a mismatch between the people and the bottleneck. “Patient capital” gets printed in the press release while the parts of the business that actually need patience — iteration, site validation, delivery ramp — are the first things speed runs over.
The company’s CEO said in June, according to the report, that the order reserve had “far exceeded our expectations”. An order book that exceeds expectations, matched to a delivery system that has not been built, does not produce sudden happiness. It produces a very quiet crack.
Reporting basis: the timeline, financial figures and allegations of unpaid wages come from the Chinese business outlet Yingzi Beiwanglu (23 October 2026) and the public filings it cites. Latest Robot News has not independently verified the wage claims, and according to that report the controlling shareholder could not be reached. Company figures for Sundry Robotics and Unitree are from their published financial disclosures. The framework-order comparison, the checklist and the analysis of incentive tempo are Latest Robot News editorial.