2026-07-29 21:01

Lead-in: Genuine robot operational data amounts to only 500,000 hours; intelligent dexterous hands are functional in just one-third of cases; systems fail upon factory transition—mass production year collides with CloudMinds-style losses and order drought

At the ICRA 2026 exhibition in Vienna, robots from over a dozen Chinese companies demonstrate tasks like folding clothes, pouring water, and tightening screws, drawing large crowds at their booths.

One visitor films a robot’s dexterous hand picking up a screw with a smartphone, while a nearby European engineer whispers to a colleague: “Is this truly fully autonomous?”

The answer to that question remains uncertain for everyone involved.

Over the past six months, RMB 46 billion has flowed into the embodied intelligence sector. If converted into hundred-yuan bills, this sum could stretch from Beijing to Shanghai and back again.

But what has it actually purchased? A burgeoning trillion-dollar industry on the cusp of explosion—or merely polished demo videos and press releases filled with fundraising news?

One: The Lessons from CloudMinds Are Still Fresh

Data from IT Juzi shows that in the first half of 2026, there were 288 funding rounds across 226 domestic companies in the embodied intelligence and robotics space, with disclosed funding exceeding RMB 46 billion.

When extending the period to July 2025 through June 2026, the figures become even more staggering: 503 funding events and over RMB 96 billion raised.

Funding is increasing, but the number of recipients is shrinking.

In the first half of the year, the top five companies—Qianxun Intelligence, Xiwang Sunrise, Xinghai Map, Ziliang Robot, and Jijia Vision—secured approximately RMB 17.1 billion, accounting for 37% of the entire industry.

The top 20 companies captured 70% (around RMB 33 billion), leaving less than 30% (about RMB 12.4 billion) for over 200 other firms. Just Qianxun Intelligence alone raised RMB 4.5 billion, completing four funding rounds within four months.

What’s worth noting is who is providing the capital. Traditional VCs remain active—Hillhouse Capital invested 13 times, Sequoia Capital 10 times.

Yet in deals exceeding RMB 1 billion, the primary investors have shifted: Baidu, ByteDance, Xiaomi, Meituan, SAIC Motor, Huachuan, and local government-affiliated investment platforms now dominate. Industrial capital and state-owned entities combined account for over 40%.

Baidu’s strategy is clear: it appears in both the RMB 1 billion Series B round of Zhifangping and the RMB 700 million Series A round of Beijing Humanoid Robotics Innovation Center.

Meituan and Didi invested in Digua Robots; SAIC Motor has invested in four embodied intelligence startups within half a year.

State capital’s approach is more direct: in transactions exceeding RMB 100 million, state participation reaches 42%. Local governments follow a three-step logic: provide funding, require factories to be built locally, and open up local manufacturing plants as the first customers.

2026 is dubbed the “Year of Mass Production” by the industry. Yujue sold over 5,500 humanoid robots last year—number one globally—with revenue surging from RMB 159 million to RMB 1.699 billion. Zhiyuan Robotics delivered its 10,000th general-purpose embodied robot in March.

But digging deeper into the numbers reveals subtle truths: securing funding does not equate to survival—the lesson is being repeatedly validated.

CloudMinds, once valued over RMB 20 billion and having raised over RMB 5.4 billion, generated only RMB 1.4 million in sales in the first seven months of 2025, suffering a net loss of RMB 84.25 million.

Two: Models Are Still Infants

While humanoid robots are selling briskly, few have actually entered factory assembly lines. The issue isn’t joint flexibility—it’s cognitive deficiency.

An industry consensus is forming: the scarcity of high-quality physical-world interaction data is the true ceiling for embodied intelligence.

Global available real-machine data totals only about 500,000 hours—far less than the text data consumed during training of large language models, which exceeds this figure by over 20,000 times. Gao Jiyang of Xinghai Map stated: “The smarter the robot becomes, the more it learns—not the cheaper it gets through mass production.”

Learning a single motion requires data that cannot be scraped from the internet; it must be gathered through repeated trials in real environments.

Thus, companies are investing heavily in data collection.

Xinghai Map launched a million-hour real-data initiative in Yizhuang; Qianxun Intelligence deployed over 300,000 data collection points nationwide; Ant Lingbo filtered out 20,000 hours from massive raw datasets, solely to train version 1.0 of its model; JD.com claims to accumulate 10 million hours within two years.

Yet results remain underwhelming. An algorithm lead interviewed by ZhiXie Dao admitted privately that after spending tens of millions collecting 100,000 hours of data, model performance improved by only 5%. Skills learned in Factory A typically fail when transferred to Factory B.

Technical roadmaps remain unresolved.

Over the past year, VLA models and world models have evolved from opposition to integration. VLA advocates direct action based on perception (“see it, do it”), whereas world models emphasize understanding physical laws before acting.

Regardless of the path, current model maturity remains in infancy. One practitioner offered a metaphor: if ultimate robotic capability scores 100 points, today’s industrial manipulators score around 50, wheeled bases 40, quadrupedal platforms 30, bipedal humanoids 15, dexterous hands 5, and associated AI capabilities only 3.

Another recurring concern: no objective benchmark exists for evaluating model quality.

Xu Huazhe of Breakout Robotics pointed out that the industry currently relies on leaderboard rankings and live demos—but ordinary users can’t experience a robot like they would a large language model.

The true evaluation metric should be: place the robot in an unfamiliar environment and observe how quickly it becomes productive.

Three: National Team Steps In to Build Infrastructure

If you map the distribution of China’s embodied intelligence companies, a clear dividing line emerges.

Beijing secured 81 funding rounds, totaling RMB 18.85 billion—40% of the national total. Qianxun, Xinghai Map, and Galaxy General are headquartered there, focusing on cognition and core architecture.

Guangdong saw 71 rounds, specializing in hardware, dexterous hands, and joint modules. Jiangsu, Zhejiang, and Shanghai combined achieved 117 rounds, primarily focused on application scenarios such as industrial coating, cleaning services, and home companionship.

The emerging structure is taking shape: Beijing supplies brains, Guangdong provides limbs, and Jiangsu-Zhejiang-Shanghai delivers workstations.

Meanwhile, inter-city competition is intensifying.

On May 8, Shenzhen Bao’an partnered with Qianhai to officially launch the “Embodied Intelligence Port”—a new industrial landmark spanning over 5 million square meters, already attracting Tencent, Galaxy General, and Luming Robotics.

Shanghai aims to deploy 100,000 humanoid robots in factories by the end of the 14th Five-Year Plan. Meanwhile, on May 1 this year, the nation’s first local regulation specifically targeting embodied intelligence robots took effect in Hangzhou.

Policy-level actions are even more substantial. On June 9, MIIT and SASAC jointly initiated the Annual Real-World Training Initiative, aiming to deploy robots in real-world industrial, service, and special applications by year-end, identifying over 100 high-value use cases and enabling deployment at scale of thousands.

The State Council Development Research Center forecasts that this market will reach RMB 400 billion by 2030 and surpass RMB 1 trillion by 2035.

Yet public funds are not easily accessible—the hidden cost of state capital involvement includes geographic lock-in, earn-out clauses, and exit restrictions.

The cautionary tale of the photovoltaic industry looms close: in 2024, 24 major PV firms collectively incurred losses exceeding RMB 28.6 billion. The current state capital participation rate in embodied intelligence mirrors that of early-stage PV development.

Exporting overseas offers another route—Yujue’s overseas revenue has long exceeded half its total.

But European markets come with European challenges: differing user habits and higher regulatory barriers.

Four: Conclusion

Returning to the afternoon when Yujue passed its IPO review, an examiner stamped a red seal on the document—the first Chinese embodied intelligence company listed on the A-share market was born.

Although capital has already pushed the sector into the “Year of Mass Production,” robots capable of stable, reliable operation—sufficient to make customers willing to pay—have yet to appear at scale.

Optimists argue: railway bubbles, internet bubbles, and new energy bubbles—all began by inflating the market, allowing real industry to emerge afterward.

Pessimists crunch the numbers: seed and angel rounds in the first half of the year totaled less than RMB 1.3 billion, accounting for only 3% of the entire sector.

A young entrepreneur without corporate backing or academic credentials may harbor the next breakthrough in his mind—but still might not even get a meeting with investors.

The collapse of CloudMinds is not an isolated case in hard-tech sectors. Every retrospective reveals similar pitfalls: overestimating technological maturity, underestimating engineering complexity, then realizing that fundraising prowess never equals survival capability.

Many companies will fall behind, disappear, and become cautionary tales in future case studies. This is not pessimism—it is the inevitable market cleansing inherent to every emerging industry.

The real answer lies neither in funding news nor in investment brochures, nor in ranking lists. It lies in the robots still running.

Can they complete a full shift? Can clients proactively place additional orders? Can foreign engineers lower their phones—not because of elegant movements, but because those movements genuinely deliver value?

Source: ZhiXie Dao