From "Thousand Units Delivered" to "Ten Thousand Units Mass Production" Embodied Intelligence Breaks Through "Dual 80%"

Financial Times Reporter Nie Yinghao

From the top-tier performers on the Spring Festival Gala stage to the quiet workers in factories, and increasingly to the “baristas” and “explainers” in public service scenes, China’s humanoid robot industry is booming. As one of the six major future industries, China’s embodied intelligence industry has achieved a leap from experimental testing to small-batch delivery, with Yushutec and Zhiyuan Robotics leading the industry with the highest global shipment volumes.

Amid the frenzy of capital pouring in to compete and expand, the embodied intelligence industry still faces core challenges: a robot capable of performing martial arts flips cannot even fold a thin piece of clothing well.

When humanoid robots can complete 80% of voice commands in 80% of unfamiliar environments, the “ChatGPT moment” for embodied intelligence will truly arrive—this industry benchmark, proposed by Yushutec founder and CEO Wang Xingxing, is inspiring countless practitioners to work tirelessly and launch into the ultimate battlefield.

From Technological Innovation

To Scene Implementation

As a representative future industry, embodied intelligence has received significant policy support at the top level, with the first full industry chain standard system released in February.

Under national policy guidance, different regions are competing with unique strategies. Beijing leverages its talent advantage to focus on “technological innovation + scene creation,” Shenzhen relies on its complete electronics supply chain to emphasize “hardware manufacturing + scene deployment,” and Hangzhou adopts a refined “city-wide integrated” spatial layout.

By 2025, the embodied intelligence industry is expected to achieve rapid development, with humanoid robot shipments already taking shape. According to IDC’s “Global Humanoid Robot Market Analysis,” global shipments are projected to reach about 18,000 units in 2025, a year-on-year increase of approximately 508%. In the global competitive landscape, Chinese companies dominate, with the top six global shipment leaders all being Chinese firms, including Yushutec and Zhiyuan Robotics forming a “dual leading” pattern. Thanks to advantages like full industry chain support and rapid engineering transformation, Chinese companies are becoming the core engine driving global growth in embodied intelligence.

The acceptance of Yushutec’s IPO application marks a new chapter, as embodied intelligence is highly sought after in capital markets. According to IT Juzi data, funding in the embodied intelligence sector in 2025 exceeded 51.1 billion yuan (3.5 times the previous year), with nearly 30 billion yuan raised in just the first three months of 2026. The capital influx has sharply raised valuation thresholds, with 13 companies now valued at over 10 billion yuan.

The Challenge of Generalization

Over the past year, the movement and operational capabilities of humanoid robots have advanced rapidly. Robots can now perform martial arts demonstrations, backflips, play piano, and even play Go. They can handle heavy lifting in factories and perform material loading tasks. However, they often struggle to hold a glass steadily, pick up a needle, or navigate around a chair. “They can beat world champions at chess but can’t open a bottle of water,” is a common market joke about embodied intelligence.

Currently, China’s embodied intelligence industry is on the cusp of large-scale production, moving from “thousand-unit deliveries” to “tens of thousands of units.” Model generalization ability remains the core bottleneck: robots can only perform specific actions according to preset programs.

What level of generalization is needed to truly realize the “ChatGPT moment” for embodied intelligence? Wang Xingxing believes that this moment means robots can, through voice or text commands, successfully complete about 80% of tasks in 80% of unfamiliar environments. For example, if a humanoid robot is taken to a completely unknown scene where it doesn’t recognize the people, and asked to “help deliver this bottle of water to someone” or “find a pen,” it can autonomously complete these tasks without prior mapping or preset programming.

In AI, there is the “Moravec paradox”: actions that are simple for humans, like folding clothes or grabbing cups, are extremely difficult for robots; conversely, tasks that are hard for humans, like complex calculations or playing chess, are easy for robots.

Tian Feng, director of the Fast and Slow Thinking Institute, explained that “walking and somersaulting” are essentially closed-loop dynamic balance and physical control problems, while “picking needles or handling fragile items” are open-loop, high-precision multi-modal perception and contact dynamics issues. The former has high fault tolerance, but the latter, with deviations as small as 0.1 Newtons, can damage eggs, cakes, or tissues.

“Breaking through fine motor skills must rely on mechanical design, highly sensitive sensor networks, dedicated low-level AI chips, and deep system-level coupling with embodied large models to truly unlock commercial potential,” Tian Feng said.

From “Body” to “Brain”

To address the generalization challenge, by 2026, the focus in humanoid robot competition is shifting from “body” to “brain,” promoting “cognition-action integration.” “Hardware determines the upper limit of robot capabilities—that is, the physical maximum—while the brain (algorithms) determines the lower limit of actual performance, namely, the ability to generalize across tasks,” Tian Feng explained.

Investments are increasingly flowing into the “brain” of robots. Since December 2025, funding in the embodied intelligence field has surpassed that of humanoid form development, with capital focusing on general embodied brains and AI motion control technologies.

Yushutec’s recent IPO prospectus shows plans to raise 2.022 billion yuan to fund research on intelligent robot models, especially the integration of embodied large models with robot motion control. Since 2026, companies like Ziyuan Robotics and Galaxy General have completed large-scale financing, with funds dedicated to R&D of embodied large models and related infrastructure.

Historically, robot operation relied on preset programs, and environmental changes could cause failures—mainly because they lack understanding of physical world laws.

He Chuan, general manager of Ningbo Embodied Intelligence Robot Innovation Center, said that the true value and industry height of robots depend on physical AI (the ability to understand the physical world), not just hardware. To push embodied intelligence toward a universal labor force era, breakthroughs are needed in spatial intelligence, physical understanding, and autonomous decision-making. Physical AI, which gives robots “physical intuition,” combined with large models driving semantic intelligence, is the key.

However, physical AI fundamentally involves replicating physical laws, which presents multiple challenges. One major issue is the lack of high-quality real-world data, especially industrial process data.

Tan Min, chief brand officer of UBTECH, explained in an interview that only through extensive real-machine training data can technological iteration be supported. For example, UBTECH collects data in real factory scenarios, trains models for specific tasks, and strives to restore the complex variations of real scenes, improving the success rate of humanoid robot tasks and ultimately enhancing their generalization level.

Manufacturers of humanoid robots are accelerating efforts to solve the “data drought” and generalization issues. “From the industry development perspective, the evolution of humanoid ‘bodies’ outpacing ‘brains’ is a temporary phase, because hardware iteration cycles are shorter, while data accumulation and model generalization require time,” said Yao Maoqing, senior vice president of Zhiyuan Robotics. “Building big brains doesn’t mean neglecting hardware; it demands higher standards for soft-hardware synergy.”

Recently, Wang Xingxing revealed at the Yabuli Forum that Yushutec’s latest plan involves full-body remote operation systems, aiming to deploy thousands, even tens of thousands, of humanoid robots by the end of this year. “By collecting 10 hours of data daily, the data problem for humanoid robots can be basically solved within the next one to two or three years.”

Supply Chain Bottlenecks

Beyond the slow development of “brains,” the supply chain bottleneck in “body” mass production remains a pressing issue. Currently, leading domestic companies like Yushutec and Zhiyuan Robotics ship over 5,000 units annually, still short of the 10,000-unit target.

“Reaching the scale of 10,000 units is a key milestone for the humanoid robot industry,” said Jiang Zheyuan, chairman of Songyan Power. “In 2026, Songyan Power aims to deliver 10,000 units of its humanoid robot ‘Xiaobu Mi.’”

Moving from prototype delivery to mass production still faces many “hard bones.” According to several senior executives in the humanoid robot manufacturing sector, an immature supply chain is a major bottleneck for industry scaling.

“The core difficulty in mass production from thousands to tens of thousands of units lies in standardizing supply chain engineering and ensuring reliability in key components like joints and sensors. Currently, these components are far from reaching the scale of the automotive industry,” Yao Maoqing told FT.

Jiang Zheyuan also acknowledged that material supply chain issues need urgent resolution. Humanoid robots involve hundreds or thousands of different parts, and a shortage of even one component can halt entire production—an issue common in early-stage mass production.

Junpu Intelligent, a “central platform” in the humanoid robot industry chain, deeply feels the difficulties of overall industry coordination: mismatched demands upstream and downstream, dispersed core component ecosystems, immature supply chains, lack of real-scale data scenarios, and limited training iterations.

“For example, there is a ‘last mile’ gap between the robot body and real industrial scenes. Manufacturers often lack deep understanding of industrial processes like precise assembly, plugging, tightening, and flexible grasping, making it difficult for robots to operate stably in factories. Additionally, demand misalignment exists: robot manufacturers focus on general architecture and basic capabilities, while factories emphasize process parameters, reliability, cycle times, and robustness, leading to low standardization,” He Chuan explained.

A typical case involved an automotive company testing embodied intelligence in assembly lines. Due to vibration from robot-line interaction and temperature fluctuations, positioning accuracy dropped sharply, causing defect rates to soar 32 times compared to laboratory conditions.

“Currently, many companies are involved in the humanoid robot upstream and downstream industries, but collaboration still has room for improvement. A truly mature industry chain would lower the overall cost of humanoid robots and accelerate their mass production,” Tan Min emphasized.

As industry competition intensifies, 2026 may become the “final exam year” for humanoid robots. Companies need to continuously deliver mass-produced and technologically advanced products. Tian Feng pointed out that 2026 will be a comprehensive contest integrating “top AI algorithms, precise manufacturing supply chains, and commercial scene deployment.” Only those with full-stack self-developed hardware and software, capable of pioneering the “data flywheel” and “business closed-loop” in real scenes, will truly establish themselves in the future trillion-dollar blue ocean. With market nurturing and technological progress, large-scale industrialization of embodied intelligence is inevitable.