What has allowed China's robotics sector to break through intense competition and stand out?

In August 2026, over 300 enterprises crowded into the exhibition hall of the World Robot Conference at Beijing Etrong International Exhibition & Convention Center, with more than 150 new products making their debut. One robot was folding clothes on site, another was giving massages to visitors. These scenes were filmed into short videos and circulated on social media. But if you pull the camera back, you will see a more thought-provoking scene: the component supply chain for the vast majority of robots in the exhibition hall, from motors, reducers to sensors, can find suppliers in some industrial parks in the Yangtze River Delta or Pearl River Delta, with a driving distance of no more than three hours. This high density of supply chains is a key to understanding the current situation of China's robotics industry.

U.S. market research firm SmartAnalyticsGlobal released a set of data in early August: in the first half of 2026, the global shipment of humanoid robots was about 19,100 units, nearly three times higher than the 5,100 units in the same period last year. Chinese manufacturers took more than 97% of the market share. Zhiyuan Robotics shipped 8,400 units, accounting for 44% of the global total; Unitree Robotics 5,900 units, accounting for 31%. The two companies together contributed about 75% of global shipments. Tesla, Figure AI and Agility Robotics were left far behind, with a huge gap.

97%. This figure is somewhat beyond expectations, but it has indeed happened. A year ago, the share of Chinese manufacturers was 84.7%, and it jumped by more than ten percentage points within a year. This is not an accidental outbreak of a single enterprise, but the simultaneous acceleration of the entire industrial chain.

The question is: Why?

AI-generated Image

The way to cut costs is more straightforward than imagined

Humanoid robots have long been confined to laboratories, and the most direct reason is their high cost. Boston Dynamics' Atlas costs hundreds of thousands of dollars per unit. Elon Musk set the ultimate target price of Optimus at $20,000, but the prototype is still far from reaching this goal. China's solution is very direct: drive down the price of components.

A joint motor that sold for 50,000 to 60,000 yuan in 2018 now costs only 500 to 600 yuan. For harmonic reducers, Japan's HarmonicDrive once dominated the market with a unit price of 2,000 to 3,000 yuan. After domestic manufacturers LeaderDrive and Laifu Harmonic started mass supply, the price dropped to 1,500 to 2,000 yuan, a decline of more than 50%. Planetary roller screws, servo systems, controllers, these components once monopolized by Japanese and German enterprises, are being overcome by domestic suppliers one by one. According to data from the Ministry of Industry and Information Technology, the localization rate of core components of humanoid robots has increased from less than 50% to more than 90%.

This means that for a humanoid robot with 40 degrees of freedom all over the body, the hardware cost of joints alone can be saved by 50,000 to 60,000 yuan. Bank of America's research believes that the cost of complete machines in China can be controlled at about 50% of similar overseas products. Unitree Robotics' G1 has a retail price of $5,600, while Tesla Optimus is still hovering in the range of $25,000 to $30,000. The price gap is an order of magnitude, so the balance of market choice will naturally tilt.

But the cost advantage is only the surface. What lies deeper is the iteration speed. The supply chain network in the Pearl River Delta allows a startup company to get samples of customized components within two weeks, while overseas enterprises may need two months. It took less than three months for Zhiyuan Robotics to roll out the 15,000th unit from the 10,000th unit. This speed is not simply achieved by working overtime, but supported by the response capability of the entire manufacturing ecosystem.

The implementation of scenarios requires working first and then optimizing. Chinese robot enterprises have a feature that they are not keen on grand narratives, but prefer to put robots into factories to make them work first.

In the first half of 2026, industrial and commercial applications have accounted for more than 70% of global humanoid robot shipments, while this proportion was only 50% a year ago. Zhiyuan's Elf G2 clusters have settled in the Nanchang factory of Longcheer Technology, operating continuously for more than 64 hours with a success rate of 99.99%. Galaxy General's Galbot S1 has entered the production line of CATL. FAW Group has built 37 typical scenario datasets around the six major links of stamping, welding, coating and final assembly. These are not demos in the laboratory, but real production lines, real operation rhythms, and real ROI calculations.

Behind this pragmatic route lies a harsh reality: humanoid robots are still not reliable enough in open environments. At the 2025 Beijing Yizhuang Robot Marathon, multiple robots fell, collided and even shut down on the track, exposing the gap between video presentation and on-site reliability completely. For standardized links, traditional industrial robots have higher reliability and efficiency; for non-standardized links with uncertainties, humanoid robots have advantages in flexibility, but will lose accuracy and stability.

Therefore, the strategy of Chinese enterprises is to first occupy scenarios that require high flexibility and relatively loose absolute accuracy requirements, such as material handling, sorting and inspection. They accumulate data in these scenarios, optimize models, and then gradually penetrate into more complex tasks. Peng Zhihui, CTO of Zhiyuan Robotics, once said: To judge whether a robot has truly stepped out of the Demo stage, we should not look at how fancy its movements are, but see whether it can stably complete complex tasks in real open scenarios and calculate clear ROI.

Bottleneck Challenges and Quiet Breakthrough

The biggest bottleneck of embodied intelligence is not hardware, but data. Large language models can crawl text on the Internet, but there is no ready dataset for robots to learn to hold an egg without breaking it or catch a water cup without dropping it. High-quality embodied data is far scarcer than text data, which is a common problem faced by the global industry. China's response is to build data collection centers.

According to Bloomberg reports, China has opened 64 data collection centers across the country, and another 20 are under construction. These facilities simulate supermarkets, assembly lines, offices, shops and home environments, allowing robots to make repeated mistakes in controlled but near-real scenarios. The Yibin Southwest Embodied Intelligence Training Center has deployed 300 robot bodies, locking in real machine data collection orders of more than 150,000 hours throughout the year. The China Unicom Embodied Intelligence Pilot Base has precipitated 5 scenarios, more than 100 atomic skills, 20,000 sets of real machine data, 5 million trajectory data, and 80TB datasets. Beijing Humanoid launched a data collection plan for 1,000 robots in real scenarios, allowing robots to collect data while working.

Morgan Stanley wrote in a report that this level of scale is the only decisive factor for ultimate success. This statement is a bit absolute, but the direction is correct. When a robot goes online, it becomes a data collection point; the data flows back to train the model, and the model becomes smarter before being deployed to more robots. After this flywheel starts to rotate, the gap will widen wider and wider.

However, it needs to be supplemented here: at present, most of the data still comes from structured or semi-structured scenarios, and the long-tail problems in the real open environment — such as pedestrians breaking in suddenly, sudden light changes, and slippery ground — are not sufficiently covered. Simulation data can reduce costs, but the effect of data collected by real bodies is still the best. The data flywheel has started to rotate, but its rotation speed and coverage still need time to be verified. Capital boom and IPO pricing moment In the summer of 2026, the capital density in the embodied intelligence track reached a peak.

Unitree Robotics landed on the Sci-Tech Innovation Board on August 19, with an issue price of 150.80 yuan, a market value of 61 billion yuan, a P/E ratio of 219 times, and about 9.78 million valid subscription households, setting a new all-time high in the history of the Sci-Tech Innovation Board. Zhiyuan Robotics launched the Hong Kong stock listing process at the end of July, with a target valuation of HK$40 billion to HK$50 billion. DeepMotion completed a Pre-IPO round of financing of nearly 200 million US dollars, with a post-investment valuation of 15 billion yuan. Lingyu Intelligence has completed 5 rounds of financing in more than a year since its establishment, and Qianting Lease has a valuation of 7 billion yuan in less than half a year since its establishment.

The logic of capital inflow is very clear: when the shipment volume jumps from the level of thousands of units to the level of tens of thousands of units, the track has entered a quantifiable stage from the story-telling stage. Unitree's listing has set a coordinate for the industry, proving that this field can generate real money valuation. But the other side of the coin is that Unitree's non-net profit deducted in the first half of 2026 decreased by 19.34% year-on-year, with a significant increase in R&D and sales expenses, showing a signal of revenue growth without profit growth. In its revenue structure, the scientific research and education scenarios account for as high as 73.6%, and industrial applications only account for 9.01%. The largest buyers are still universities and scientific research institutions, not factories.

This means that mass production data can be generated, and financing enthusiasm can be boosted, but the maturity of commercialization can only be achieved through time of polishing. 2026 can be the first year of mass production, but not necessarily the first year of profitability. At present, the biggest hidden concern is that the "body" is strong, but the "brain" is still catching up. If we draw a capability map for China's robotics industry, the "body" segment — that is, hardware manufacturing, supply chain integration, and large-scale mass production — has indeed been at the forefront of the world. But the "brain" segment, that is, AI decision-making, generalization capability, and high-end software tool chains, still has obvious shortcomings.

For example, a large number of Chinese enterprises are still using NVIDIA's IsaacSim, Cosmos, GR00T and Orin ecosystems for training and simulation. High-end computing power, EDA tools, advanced processes and some software tool chains rely on Western supply, which is similar to the structure of the AI industry. End-side chips currently only support models with tens of billions of parameters, and the deployment solutions for larger models are not yet mature.

Dexterous hands are another key bottleneck. In the past few years, the public was easily attracted by the robot's movements of running, jumping and somersaulting, but underestimated the importance of hands. The vast majority of human precision operations rely on hands, and the durability and generalization capability of industrial-level dexterous hands have not been fully proven worldwide. Figure AI piloted in BMW's factory for 11 months, handling more than 90,000 parts, but the forearm is still the part with the highest failure rate.

The deeper problem is that when robots move from Demo to deployment, engineering indicators such as maintenance cost, batch consistency, fenceless safety, and long-term reliability are more important than somersaults on the stage. The Unitree team revealed that the somersault movement for the Spring Festival Gala was iterated for more than 300 versions, and the synchronization error of 26 robots was controlled within 0.1 seconds. But all of this happened in a highly controlled stage environment. The real test takes place in scenarios such as factory night shifts, continuous operation, and unattended operation.

The essence of breakthrough is a systematic engineering. It is one-sided to attribute the breakthrough of China's robotics industry to a single factor. It is not achieved by a certain technological breakthrough, nor by piling up policy subsidies. Policies provide top-level design and trial-and-error space, supply chains provide cost advantages and iteration speed, scenarios provide implementation verification and data backflow, and capital provides amplifiers and valuation anchors. These lines are intertwined to form the current situation.

In 2026, the market size of China's embodied intelligence is expected to reach 6 billion yuan, and the output of humanoid robot complete machines is expected to exceed 100,000 units. Behind these figures, a large manufacturing country has migrated the supply chain organization capabilities accumulated in the fields of consumer electronics and new energy vehicles in the past 30 years to the robotics track. The methodology of cost reduction, mass production, iteration and implementation has been verified twice, and now it is the third time. But there is still a long way to go before the real universal robot. Only when robots continuously run into production lines, stores, and real workflows, and can stably complete complex tasks in open environments, can the industry move from concept verification to deployment stage. Before that, keeping a little caution and retaining some expectations may be a more appropriate attitude.

In the exhibition hall of the World Robot Conference, those robots that fold clothes and give massages will eventually enter the daily lives of more people. But when and in what way depends on how fast the data flywheel rotates in the next few years, how low the supply chain can push the cost, and whether those engineering problems hidden in joints, reducers and algorithms can be overcome one by one.

This breakthrough has only just begun.

This article is from "Sice Think Tank", authorized for release by 36Kr.