Produced by | RoboIsland

At ICRA 2026 in Vienna, robots from a dozen Chinese companies are demonstrating tasks like folding clothes, pouring water, and screwing in bolts, drawing crowds to their booths.

Someone holds up their phone to capture a robot deftly picking up a screw with its nimble hand, while a European engineer nearby whispers to a colleague, "Is this really fully autonomous?"

No one dares to say for certain.

Over the past six months, 46 billion yuan has poured into the embodied AI sector. If converted into 100-yuan bills laid end-to-end, this sum could stretch from Beijing to Shanghai and back again.

But the question on everyone's mind is whether this investment will fuel a trillion-yuan industry or simply result in a collection of polished demo videos and funding press releases.

I. The Lesson of CloudMinds Looms Large

Data from IT Juzi shows that in the first half of 2026, there were 288 financing events in China's embodied AI and robotics sector, involving 226 companies, with disclosed funding exceeding 46 billion yuan.

Extending the timeframe to July 2025-June 2026, the numbers soar: 503 financings and over 96 billion yuan.

While capital inflows are rising, the number of recipients is shrinking.

In the first half, the top five companies—Qianxun Intelligence, Sunrise Robotics, Xinghaitu, Autonomous Robotics, and Jijia Vision—raised approximately 17.1 billion yuan, accounting for 37% of the industry total.

The top 20 firms secured 70% (about 33 billion yuan), leaving the remaining 200+ companies to split less than 30% (around 12.4 billion yuan). Qianxun Intelligence alone raised 4.5 billion yuan across four rounds in four months.

Notably, the sources of funding are shifting. Traditional VCs remain active, with Hillhouse Capital making 13 investments and Sequoia Capital 10.

However, for deals exceeding 1 billion yuan, the main players are now Baidu, ByteDance, Xiaomi, Meituan, SAIC, Inovance, and government-backed investment platforms. Industrial capital and state-owned funds now account for over 40%.

Baidu's strategy is clear: it participated in both Intelligent Square's Series B round (1 billion yuan) and the Beijing Humanoid Robot Innovation Center's Series A round (700 million yuan). Meituan and Didi invested in Digua Robotics, while SAIC backed four embodied AI firms within six months.

State-owned capital's entry is even more direct: in deals exceeding several hundred million yuan, state participation reached 42%. Local governments follow a three-step logic: provide funding, require companies to build factories locally, and secure them as anchor customers.

2026 is hailed as the year of mass production. Unitree sold over 5,500 humanoid robots last year, ranking first globally, with revenue surging from 159 million to 1.699 billion yuan. Zhiyuan rolled out its 10,000th general-purpose embodied robot in March.

Yet, a closer look reveals nuances. Securing funding does not guarantee survival, a lesson repeatedly validated.

CloudMinds, once valued at over 20 billion yuan with over 5.4 billion yuan in funding, sold just 1.4 million yuan worth of products in the first seven months of 2025, incurring an 84.25 million yuan net loss.

II. The Model is Still a Child

While humanoid robots sell briskly, few make it onto factory floors. The issue lies not in inflexible joints but in insufficient intelligence.

A consensus is emerging: the scarcity of high-quality physical interaction data is the true bottleneck for embodied AI.

Globally, usable real-world robot data totals about 500,000 hours—a mere 0.05% of the text data consumed in training large language models. Gao Jiyang of Xinghaitu observes, "Robots grow smarter by learning more, not by being built cheaper in larger quantities."

But acquiring data for a single action requires real-world trials, not web scraping.

Companies are now investing heavily in data collection. Xinghaitu launched a million-hour real-data initiative in Yizhuang; Qianxun Intelligence deployed over 300,000 collection points nationwide; Ant Linkwave filtered 20,000 hours from massive raw data to train its 1.0 model; JD aims to accumulate 10 million hours within two years.

The results are underwhelming. An algorithm leader interviewed by RoboIsland admitted that spending tens of millions on 100,000 hours of data improved model capability by just 5%. Skills learned in Factory A often fail in Factory B.

Technical pathways remain unsettled.

Over the past year, VLA (Vision-Language-Action) models and world models have moved from rivalry to convergence. VLA advocates "seeing and acting," while world models prioritize understanding physical laws before acting.

Yet, regardless of approach, model maturity remains in its infancy. An industry insider compares it to a scoring system: if ultimate robot capability is 100 points, current industrial arms score ~50, wheeled bases ~40, quadrupeds ~30, bipedal humanoids ~15, dexterous hands ~5, and supporting AI ~3.

Another recurring issue: no objective metrics exist to evaluate model performance.

Xu Huazhe of Poke Robotics notes that the industry now focuses on leaderboard rankings and demos, but ordinary users cannot test robots as easily as large language models. A valid benchmark would be deploying a robot in a new environment and measuring how quickly it adapts.

III. The State Team Enters to Build Infrastructure

Mapping China's embodied AI companies reveals a clear geographical divide.

Beijing secured 81 financings totaling 18.85 billion yuan, accounting for 40% nationally. Qianxun, Xinghaitu, and Galaxy Universal are based here, focusing on "brains" and robot bodies.

Guangdong attracted 71 deals, specializing in hardware, dexterous hands, and joint modules. The Jiangsu-Zhejiang-Shanghai region accounted for 117 financings, focusing on application scenarios like industrial spraying, cleaning services, and companionship.

The industry landscape has taken shape: Beijing supplies intelligence, Guangdong provides limbs, and Jiangsu-Zhejiang-Shanghai offers workstations.

Meanwhile, intercity competition intensifies.

On May 8, Shenzhen's Bao'an District and Qianhai officially unveiled the "Embodied AI Port" industrial landmark, spanning over 5 million square meters, attracting Tencent, Galaxy Universal, and Luming Robotics.

Shanghai aims to deploy 100,000 humanoid robots in factories by 2030. On May 1, Hangzhou implemented China's first local regulation focused on embodied AI robots.

Policy moves are bolder. On June 9, the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) launched an annual real-world training initiative, targeting robot deployment in industrial, service, and specialty sectors by year-end. The goal is to create 100+ high-value application scenarios and enable 10000 units level (ten-thousand-unit-scale) deployments.

The State Council Development Research Center predicts the market will reach 400 billion yuan by 2030 and 1 trillion yuan by 2035.

However, state funding comes with strings attached: geographical lock-in, performance guarantees, and exit restrictions.

The solar industry's cautionary tale looms: in 2024, 24 major solar firms collectively lost over 28.6 billion yuan. The current state participation rate in embodied AI mirrors the solar sector's early days.

Going global is another path. Unitree derives over half its revenue from overseas markets.

Yet, European markets present unique challenges: differing user habits and stricter regulations.

IV. Conclusion

Returning to the afternoon of Unitree's IPO approval, as the CSRC official stamped the red seal on the documents, China's first A-share-listed embodied AI company was born.

While capital has propelled the sector to its "year of mass production," robots capable of consistently performing tasks and winning repeat orders have yet to emerge in bulk.

Optimists argue that every major tech boom—railroads, the internet, new energy—first heated the market before true industrialization followed.

Pessimists crunch numbers: seed and angel rounds in the first half totaled less than 1.3 billion yuan, just 3% of the sector's total.

A young entrepreneur without corporate backing or academic titles, even armed with the next breakthrough idea, might struggle to gain investor access.

The collapse of CloudMinds is not an isolated incident in the hard tech sector. Each retrospective reveals familiar pitfalls: overestimating technological maturity, underestimating engineering challenges, and discovering that fundraising prowess ≠ survivability.

Many firms will fall by the wayside, becoming case studies in future retrospectives. This is not pessimism but the inevitable consolidation phase of every emerging industry.

The sector's true answers lie not in funding announcements, investment documents, or leaderboard rankings but in the robots still in operation.

Whether they can complete full shifts, secure repeat orders, and convince skeptical foreign engineers to look up from their phones depends not on flashy moves but on their practical utility.