Humanoid robots are racing towards IPO.

In the past month, the capital market has been in turmoil, with billions of RMB pouring into the embodied intelligence industry.

Unitree Technology officially started the listing guidance on July 18th and began to sprint for an IPO. It had only been two months since Unitree Technology completed its shareholding reform. A month ago, on June 19th, Unitree Technology announced the completion of a strategic financing round of Series C+. According to estimates by ITjuzi, the company's valuation reached 13 billion RMB after this round of financing.

On July 8th, Zhipu Robotics, which is in the same first - tier as Unitree Technology, dropped a bombshell by investing at least 2.1 billion RMB to acquire 63.62% of the shares of Shangwei New Materials, a listed company on the Science and Technology Innovation Board of the A - share market. Although Zhipu Robotics claims that this acquisition does not involve the restructuring and listing of its business and major assets, spending more than two billion RMB to acquire a company unrelated to the robotics industry seems to be self - deceiving if there are no other motives. Wang Chao, the founder of Lanqiao Capital, called this move by Zhipu Robotics a "quasi - backdoor listing" in a conversation with LatePost, clearly showing its ambition to get closer to the capital market.

The IPO race between these two leading companies reflects the targeted support from the authorities for humanoid robots, with clear intentions. The opening of the "gate" has also driven the linkage between the primary and secondary markets.

In recent days, Hangzhou DeepClouds, one of the "Six Little Dragons of Hangzhou", announced the completion of a new round of financing of nearly 500 million RMB. Tashii announced the completion of a Series Angel + financing of $122 million led by Meituan's strategic investment arm.

It is worth noting that this is not the first time Meituan has appeared as an investor in the embodied intelligence field. According to incomplete statistics, Meituan has a stake in many embodied intelligence companies, including Unitree, Zhipu, and Xinghaitu and Independent Variable Robotics, which will be mentioned later. It can be said that Meituan supports half of the embodied intelligence industry.

JD.com wants to take over the other half.

On July 21st, three robotics companies, Zhongqing Robotics, Qianxun Intelligence, and Zhujie Dynamics, simultaneously announced that they had received a new round of financing led by JD.com. The financing scales of the first two companies reached nearly one billion RMB and nearly 600 million RMB respectively. JD.com seems to be in a head - to - head competition with Meituan, and the battlefront has spread from the "food delivery war" to the embodied intelligence industry.

Beijing Xiaoyu Zhizao completed a Series A+ financing of hundreds of millions of RMB led by Didi. This is Didi's first investment in the embodied intelligence field.

In addition, Xingdong Jiyuan, a Tsinghua - affiliated embodied intelligence company, completed a Series A financing of nearly 500 million RMB. Geek+ Technology, which focuses on warehouse robots and intelligent logistics, was listed on the main board of the Hong Kong Stock Exchange, becoming the "world's first listed company in AMR warehouse robots". Xinghaitu completed two consecutive strategic financing rounds of Series A4 and A5, with a total financing amount of over $100 million. Many other companies, such as Tusu Technology and Kuawei Intelligence, also completed financing recently. Wang Qian, the founder and CEO of Independent Variable Robotics, even said, "We completed three rounds of financing in the first half of the year."

The numerous financings that have occurred in the past month are dazzling, reflecting the capital market's expectations for the future industry of embodied intelligence. Many companies are flocking to this field, hoping to make early arrangements in this blue - ocean market. However, in Wang Qian's view, even with such a large - scale financing, the speed is still not fast enough and the scale is not large enough. When interviewed by a reporter from Jiemian News, he said, "Embodied intelligence is a major field where China can compete with the United States on an equal footing. However, compared with American companies, Chinese companies have an order - of - magnitude gap in both financing scale and company valuation."

However, the upsurge has begun, and this is a good start.

01 Where Does the Confidence Come From?

Embodied AI and robots are not new concepts. In 1950, Turing proposed two development paths for artificial intelligence in his paper: one is abstract intelligence, which relies on algorithms and symbolic reasoning; the other is embodied interaction, which equips machines with sensory devices so that they can learn skills through sensory experiences and interact with the environment.

However, it is only until now that embodied intelligence has been mentioned so frequently. Many startups have resolutely entered the field of building robots and successfully attracted a large amount of capital. So the question is, after the PC era and the mobile phone era, why has an industry that was not previously given much attention suddenly taken off and become so popular?

This involves two key paradigm shifts.

The first paradigm shift occurred in the robot's body. It is a transformation from the hydraulic - drive route to the electric - drive route.

Wang Xingxing, the founder of Unitree, once said that he realized as early as before 2013 that the hydraulic solution could not be commercialized. The reason is simple: it consists of precision mechanical parts, and once precision mechanical parts are involved, the cost will be high. Moreover, all hydraulic systems leak oil, and even household cars rarely use hydraulic systems anymore.

The old giants that have been on the hydraulic technology route for decades have made little progress. In 2018, Japan's ASIMO robot announced the cessation of its R & D, which plunged the global humanoid robot industry into a trough. It was not until Elon Musk established the electric - drive route that it became a crucial technological turning point that saved the global robot industry.

The large - scale simplification of parts brought about by electric drive has enabled economies of scale. Even the hydraulic camp, which has accumulated technological barriers over half a century, has been overtaken by new small giants overnight. Boston Dynamics' annual sales are now only one - tenth of Unitree's and may gradually fade out of the historical stage.

Therefore, a huge dividend of electric drive is that it has opened a window for late - comers, from small startups to large smart home appliance giants, automobile manufacturers, and Internet giants, to catch up. All players, big and small, are now on the same starting line.

This also benefits from the mature industrial chain of new - energy electric vehicles. The motors have become smaller in size but larger in torque, and the energy density of batteries has also been continuously increasing. Transferring these technologies to the robot industry can improve the robot's mobility and endurance. Like a snowball, high - performance and low - cost components are developed one by one, gradually approaching the industry's inflection point.

The second paradigm shift occurred in the robot's "brain". It is a transformation brought about by large - model technology.

Wu Changzheng, the president of Magic Atom, a robotics company, said that the industry boomed in 2023 and 2024, and the emergence of ChatGPT 3.5 was crucial. The development of large - model technology has brought a new paradigm for improving the intelligence level of robots. The generative AI has shown a thousand - fold or even ten - thousand - fold increase in capabilities, making us realize that there is infinite potential in the combination of large models and robots. In terms of environmental perception, understanding, reasoning, decision - making, and task planning, robots will have a new paradigm. Under this new paradigm, the capabilities of robots will increase by a thousand - fold or even ten - thousand - fold.

The emergence of large models has subverted the traditional impression that robots can only be "remote - controlled toys". The complex - scenario understanding ability and long - sequence task - planning ability demonstrated by large models are not possessed by traditional robots. In the past, the task sequences of robots were pre - written by humans. However, through continuous training and learning, large models can enable robots to understand the surrounding environment and autonomously plan tasks based on the environment and target tasks, just like humans with a brain.

Of course, this does not mean that existing large models can be directly stuffed into robots. Large models interact through language, and language can be used as a medium for human - machine interaction, but it cannot be regarded as the intelligence that robots possess. Wang He, the founder and CTO of Galaxy Universal, said, "The essence of intelligence is the ability to make a corresponding reaction to a situation." For robots, it is crucial to interact with the surrounding environment through vision, hearing, and touch and have the ability to understand the current environment and take actions.

Therefore, an embodied large model suitable for robots has become an industry consensus. Although there are still technological challenges to overcome, the direction is clear, and the door to progress has been opened. We are now just waiting for a GPT - 4 moment for the robot industry.

However, the road ahead is still long and winding.

02 The Obstacles Remain

At present, embodied intelligence is still far from its ideal state. The current embodied intelligence market is like the new - energy vehicle market in the 2010s, in the early stage of the industry's wild development. The technology is not yet mature, and it is still a long way from commercialization.

There is a gap between the story that embodied intelligence wants to tell and the current reality.

First of all, the training of embodied intelligence requires a large amount of interaction data as support. Xie Junyuan, the person - in - charge of the embodied intelligence project at Qianxun Intelligence, said, "Data is currently the biggest challenge. Many problems will automatically disappear when the data volume increases."

Currently, the main sources of data collection are divided into two types: real - world data and simulation data. Real - world data is obtained through the actual operation of robots to get feedback during their interaction with the real physical world, mainly through remote operation and motion capture. Simulation data is generated by rendering virtual environments and simulating the interaction between robots and objects.

Since it involves real - machine interaction, the quality of real - world data collection is the highest, but its shortcomings are also obvious: the data collection workload is large and the cost is high. Tongji Zihao, an ecological partner of Gaoqing Electromechanical, mentioned, "I used a master - slave robotic arm to teach a robot to catch crayfish. After repeating it a hundred times, my hand was very tired." Wang He also pointed out that the cost of hiring people to collect real - world data through remote operation is very high. "The monthly cost of collecting data from tens of thousands of robots ranges from hundreds of millions to billions of RMB."

Simulation data can be used for thousands of training sessions, but there is a gap between the virtual world and the real world. The physical laws and robot perception in the simulation environment deviate from those in the real world, and the data obtained through simulation may be ineffective in reality. Tongji Zihao said, "The current mainstream solution is a hybrid training of'simulation + a small amount of real - world data', but how to narrow the'simulation - reality gap' is still the core problem."

The lack of generalization means a low task - success rate. Coupled with the high price, humanoid robots are difficult to enter households in the short term and become consumer - grade products.

Gao Yang, the co - founder of Qianxun Intelligence, divided the stages of embodied intelligence into L0 - L5 in a conversation with LatePost: L0 is non - intelligent industrial robots; L1 is single - task intelligence; L2 is multi - task intelligence in a single scenario; L3 is the ability to complete 70% - 80% of human tasks in a single scenario; L4 is the ability to complete 100% of tasks in a single scenario; L5 is all - around ability across scenarios. "The entire industry is currently on the way from L1 to L2." It will still take a long time for humanoid robot companies to occupy the consumer market.

Gao Yang said, "The main challenge in the improvement from L2 to L3 is generalization. It is difficult to collect all task data, so generalization is needed to enable robots to draw inferences from one instance." Generalization means that an intelligent agent can effectively transfer and adapt the skills it has learned in a single environment, task, or entity to previously unseen environments, tasks, or entities in the real, open, and dynamically changing physical world. This ability to adapt for a lifetime in an open world is the key bottleneck and core goal for embodied intelligence to finally become practical.

Currently, most robots can only work in highly controlled environments. Wang Xingxing once told Zhou Tian Finance, "The current problem is that a task can be completed, but if the scenario and task change slightly, the success rate will plummet." This is the problem caused by the lack of generalization. The fancy moves that robot companies have come up with, such as dancing or backflips, are essentially fixed actions or remote - controlled operations and have not fundamentally differed from traditional robots. This is also the reason why robots cannot currently enter general scenarios.

A senior investment professional said, "Currently, the largest market for pure humanoid robots is research institutions, which need human - like configurations for whole - machine control research. Other scenarios are just gimmicks - mall exhibitions, advertising shoots, and they can even be rented for one or two uses, indicating that the demand is not long - term and high - frequency."

Wang Xingxing, the founder of Unitree Technology, announced at the Summer Davos Forum a month ago that Unitree's annual revenue had exceeded one billion RMB, warming up for the company's listing. According to the statistics of the Silicon - based Laboratory, scientific research universities account for the majority of Unitree's orders, and the rest are purchased by central state - owned enterprises, government agencies, etc.

On July 11th, Unitree Technology and Zhipu Robotics jointly won the largest order for domestic humanoid robot companies so far - a bipedal robot OEM service procurement project with a total budget of $124.05 million. This order is from China Mobile (Hangzhou) Information Technology Co., Ltd.

From the above, we can see the current commercialization bottlenecks of humanoid robot companies. Under the influence of various unfavorable factors, the deployment scenarios of robots are limited to a narrow range, and more often it is for experimental and scientific research needs. The revenue sources of robot companies cannot cover the broad consumer market in the short term.

Humanoid robot companies still have a long way to go to get closer to consumers. However, the high premiums given by capital are mainly for the possibility of consumer - grade products.

03 Where Is the Future Heading?

The influx of funds indicates that entrepreneurs and investors are very optimistic about the future.

Wang Xingxing asserted, "The technological progress at the end of 2025 will reach a new level, which will give a greater boost to the global industry. It is no problem for the popularity and market size to increase tenfold."

Currently, there are two mainstream technical routes towards embodied intelligence: one is the end - to - end model, and the other is the hierarchical decision - making model. The end - to - end route directly maps raw sensor inputs such as camera images, lidar data, and tactile signals to the robot's action outputs without relying on artificially designed intermediate modules. It completes the entire process from perception to decision - making through a single model.

The core of the hierarchical decision - making model is to decompose tasks into multiple levels, with each level focusing on specific sub - functions. The links from perception to execution are independent, and each module can be optimized separately or use a hybrid technology solution.

In fact, the end - to - end model has become the consensus of most practitioners. This is the solution chosen by Tesla, and many well - known domestic embodied intelligence startups, such as