EN / 中文

How Humanoid Robot Vendors Sell Machines Before Generalization Is Ready

by hangjiashuojiqiren·November 19, 2026

Hangjia Shuo Robotics

At the APEC Business Leaders' Summit, UBTECH founder Zhou Jian stated that a general-purpose humanoid robot with strong generalization capabilities "still requires some time to develop given today's hardware and large model standards."

On the same day, Unitree CEO Wang Xingxing also explicitly pointed out: The primary pain point in the robotics industry is insufficient model generalization.

Regarding the same issue, CEOs from two different companies reached the same conclusion: Generalization has not yet arrived. So before that happens, how do robotics companies sell their machines and retain customers?

[Hangjia Shuo Robotics] attempts to review the transactions already completed in the industry.

01. The CNY 760,000 Robot Is Not Just About Buying Hardware

In 2025, UBTECH delivered 1,079 full-size humanoid robots, generating revenue of CNY821 million, with an average price of approximately CNY 760,000 per unit.

According to UBTECH's own disclosed metrics, its robots achieve material handling efficiency in industrial scenarios at about 30% to 40% of human levels, with a target to increase this to 50% by 2026 and break through 80% by 2027.

With efficiency only half that of humans and prices two to three times higher than comparable cars, it indicates that customers are not buying efficiency. Media reports put it more directly: UBTECH is not selling simple hardware, but a complete solution including the robot body, software system, scenario adaptation, and continuous operation and maintenance.

In the Zeekr 5G Smart Factory, the Walker S2 participated in battery module assembly, shortening the mass production cycle by 15%—what customers bought was precisely the fact that "this specific process could be reliably handled."

02. Why Processes Sell Better Than Robots

Because before the economics of general models balance out, processes are the smallest unit that can be clearly calculated.

A single model covering all scenarios is difficult, takes a long time, and has a high failure rate; however, a single process—moving a bin from A to B, inserting a harness into a port—has clear task boundaries, quantifiable success, and fixed scenarios. The missing part of general capability is compensated for by engineers.

UBTECH's approach is to lock robots into processes, then build outward with a three-layer structure of "hardware platform + local scenarios + local research capabilities." The customer list confirms this structure: Texas Instruments, Airbus, Honda Trading, Hitachi Group—the latter two being partners mentioned by Zhou Jian at APEC.

03. Half of the Orders Are Not Sold to Factories

There is now an unavoidable fact in the industry.

In 2025, UBTECH signed cumulative contracts worth nearly CNY 1.4 billion. From October to December, it signed at least six orders from data collection centers totaling approximately CNY 829 million, accounting for about 59% of the annual contract value. The clients were government-led data collection and testing centers in Guangxi Fangchenggang, Sichuan Zigong, and Huizhou Greater Bay Area.

The robots purchased back are primarily tasked with "being trained," not "going to work." Contract values do not equal recognized revenue, so this figure cannot be directly counted towards sales.

This is not just a problem for UBTECH alone. IDC data shows that in 2025, commercial performances, scientific research, and data collection accounted for 78.4% of humanoid robot shipments; The Economist estimated in August this year that up to 70% (about 13,000 units) of humanoid robots made in China in the first half of 2026 flowed to data collection centers.

Therefore, "selling processes" cannot be generalized: selling processes to factories is commercialization, while selling processes to data collection centers is a transitional form during the industry incubation period. Both represent real deliveries, but their reference value for order sustainability differs by orders of magnitude.

04. Rubber in Thailand and Production Lines in Japan

At the Thailand-China Cooperation Expo, Thai Prime Minister Anutin Charnvirakul personally visited the UBTECH booth, and the Deputy Prime Minister and Minister of Higher Education also inspected the "Thai Local Industry Adaptation Solution." During the expo, UBTECH signed a memorandum with the Institute of Field Robotics (FIBO, KMUTT) (King Mongkut's University of Technology Thonburi - Field Robotics Institute): opening underlying data interfaces and development toolchains for Walker S2 and Cruzr Y1, allowing the FIBO team to deeply integrate self-developed algorithms, industry models, and robot bodies. They jointly established a "Humanoid Robot Co-research and Co-creation Center," focusing on specific operational algorithms for Thailand's automotive, electronics, and new energy sectors. Thailand's "National Artificial Intelligence Development Plan (2025-2027)" plans to invest THB 25 billion to promote manufacturing intelligence.

Rubber cutting is representative of these "locally unique processes"—almost non-existent in China except Hainan, which precisely makes it a local development opportunity.

Japan follows a different line, closer to traditional industrial channels. In April, Honda Trading announced a strategic partnership with UBTECH subsidiary UQI, exploring automotive manufacturing and warehousing logistics scenarios; in May, Hitachi officially announced cooperation, introducing Walker S2 to manufacturing sites for job testing. During the China International Import Expo, Walker S2 demonstrated real material handling processes at Hitachi's factory. Hitachi treated its own factory as the "Customer Zero," replicating the solution to other clients after validation. Japan Airlines introduced UBTECH equipment at Haneda Airport for ground service experiments.

Channels have also changed. UBTECH's exclusive agent in Japan bundles RaaS (Robot-as-a-Service) and physical scenario data services into solutions—no longer selling one-off devices, but deployment, continuous service, and real-world scenario data feedback.

05. The Cost of This Business

The ceiling of process-based approaches is clear.

The average price of CNY 760,000 includes a package of the body, system, adaptation, and O&M, while efficiency is only 30% to 40% of humans. Guotai Haitong Securities judges that expansion to more processes can only be discussed after generalization improves; to achieve payback within two years, the selling price must drop to the CNY 100,000 level while matching human efficiency. Neither threshold has been crossed yet.

Zhou Jian said at the meeting, "Eventually differentiation will form, but they will also merge." The deeper the process specialization, the harder it is to step away when general capabilities arrive—customized development investments, accumulated scenario data, and tuned process parameters are assets, but also burdens.

So whether to "sell robots or sell processes" may not be a question of choice. Process-based approaches might not just be a transitional solution before generalization; they might be a long-term form of humanoid robot implementation, much like industrial software never waited for a universal computer.

What truly needs watching next is whether these process-customized robots can hand over the accumulated process adaptations when general models mature.

This article is compiled based on public online information, for reference only, and does not constitute investment advice.