Embodied intelligence is climbing over three major hurdles.

In the past decade or so, the tech circle has never been short of "trending topics".

VR glasses, virtual currency blockchain, and the metaverse. Which of these concepts didn't once seem to have boundless prospects, only to fizzle out in the end?

So, with the sudden popularity of embodied intelligence this year, many people's first reaction was: Here we go again?

01

Embodied Intelligence Has Many "Predecessors"

Indeed, like many of its "predecessors", embodied intelligence exhibits many characteristics of a trending topic.

Why is that?

Besides the well - known event of "Spring Festival Gala robots dancing", another representative event is that this year, it was included in the "Government Work Report" for the first time: "Establish a growth mechanism for investment in future industries, and cultivate future industries such as biomanufacturing, quantum technology, embodied intelligence, and 6G."

Embodied intelligence has, in fact, become an important direction for China to cultivate future industries and has also become a "key area" in the global technology competition and cooperation.

In just the first half of 2025, capital invested over 20 billion yuan, with more than 130 financing events. Tech giants like OpenAI, NVIDIA, Google, Huawei, and Alibaba have all entered the fray to grab this "hot cake".

Many people started to express pessimism.

For example, Zhu Xiaohu, the founder of GSR Ventures, directly stated that the commercial path of embodied intelligence is unclear and withdrew from the humanoid robot companies he had invested in. Fu Sheng, the CEO of Cheetah Mobile, also firmly believes that it is too difficult to implement humanoid robots and their applications are too far off. It will take at least five years, or even longer, for them to be put into use.

Even international investment banks have taken notice of Unitree Technology. After conducting a research on the company on February 27 after the Spring Festival, they released a report stating that humanoid robots still have a long way to go before they can be truly put into production and daily use, and the market's expectations for this technology may be too far - reaching.

It seems that it may follow in the footsteps of its predecessors, like concepts such as virtual currency, blockchain, AR glasses, driverless technology, and the metaverse, with the promise of realization being a long way off.

02

Double Strangulation of Technology and Cost

Difficulty in Realizing Economic Value

This concern is not unfounded.

Firstly, there are many bottlenecks in technology, such as computing power, battery capacity, and the balance between "brainpower" and physical ability.

The concept of embodied intelligence was proposed as early as 1950. Despite the development of semiconductors and the progress of programming technology, relevant startup companies have never come up with any usable products.

It was not until the emergence of concepts like Industry 4.0 that the relevant industrial chain began to develop to some extent. However, even with an algorithm with a 99% success rate in Isaac Gym, when applied to a real robotic arm, the task success rate may drop to 30%. The AI generalization ability of embodied intelligence both at home and abroad shows obvious deficiencies, and the task failure rate in non - preset scenarios (such as a messy home environment) is generally over 10%.

It was not until the emergence of concepts like Industry 4.0 that the relevant industrial chain began to develop to some extent.

But most of the development is concentrated in areas such as fixed robots, wheeled robots, quadruped robots, and tracked robots, and the tasks they can perform are simple and single.

Then came the emergence of artificial intelligence.

The surface of artificial intelligence is the algorithm, and the core is computing power. The large - scale model is an upgraded form that combines them. It can be said that future embodied intelligence with more thinking ability must be driven by large - scale models.

However, this will also face some new situations - the "brain capacity" of robots is too small. It is not easy to install a multi - modal large - scale model with hundreds of billions of parameters in a robot's "brain".

Due to the complex and changeable working scenarios of robots, the option of "cloud services", which has high latency and is unstable, is almost ruled out. So robots also need to consider the bottleneck of battery technology itself. When the battery capacity of the robot's body is limited, how to strike a balance between high power consumption and high "brainpower"? Humans took hundreds of thousands or millions of years of evolution to find an optimal balance, but robots clearly won't have that much time to evolve.

Secondly, the cost of embodied robots remains high, and there is a double - edged game between capital and cash flow.

Take "Boston Dynamics", a pioneer in the industry, as an example. It dominates the quadruped robot field, but it has changed ownership three times in the past decade, from Google to SoftBank and then to Hyundai. The reason for each change of ownership is written in the red ink of the financial statements: The cost of an Atlas experimental machine is 2 million US dollars - equivalent to a school - district apartment in the Third Ring Road of Beijing, but it can only do parkour in obstacle courses and is even afraid of spilling a cup of coffee when handing it over.

Domestic embodied intelligence companies can relate to this.

ZhiYuan Robotics rushed to submit its prospectus just two years after its establishment, simply because its cash flow was almost exhausted. In 2023, ZhiYuan Robotics lost 870 million yuan, and the cash on its books was only enough to last for 18 more months. If it didn't go public to "replenish blood", it would have to put the robotic arms in the laboratory on Xianyu.

Similarly, for Yunji Technology, a leading service robot company, its revenue increased from 161 million yuan in 2022 to 245 million yuan in 2024, with a compound annual growth rate of 23.2%. However, the losses during the same period were as high as 365 million yuan, 264 million yuan, and 185 million yuan respectively. Geek+ Technology, a leading mobile robot company closely related to embodied intelligence, had an even greater loss. In 2024, its revenue was 2.409 billion yuan, but it lost 832 million yuan. It stated in its prospectus: "We prioritize business expansion and innovation and focus on long - term value creation rather than short - term financial returns."

Putting aside the training and labor costs for now, just the hard expenses are a money - gobbling monster.

For example, the three major hardware components - reducers, servo motors, and controllers - account for 60% to 70% of the total cost of the machine. For every 0.01 - millimeter increase in precision, the budget has to be increased by an order of magnitude.

The algorithm is also not cheap. Huawei Cloud once calculated that to run a large - scale model with 1 billion parameters on a robot, it would take 10,000 A100 graphics cards working non - stop for a month, and just the electricity cost would be extremely high.

Then, there is a severe lack of scenarios for embodied robots. In terms of both work price and flexibility, they are difficult to compete with humans.

Many people may have fantasized about the scenarios in science - fiction movies or games, where robots can think and work like humans. But in fact, embodied intelligence will be difficult to meet such fantasies for a decade or two, or even longer.

Currently, almost no robot - making company will compare its products with humans in public. After all, in terms of "working", embodied robots are also typical of being "unable to do high - end jobs and unwilling to do low - end jobs".

Each G1 robot of Unitree Technology requires manual adjustment of the joint motors, and the monthly production capacity is only 300 units. The hardware gross profit margin of Boston Dynamics' Atlas is only 18%. Selling one unit results in a loss equivalent to half a unit, almost turning high - tech into "charity".

(Boston Atlas Robot)

70% of the robots produced by ZhiYuan Robotics are used in industrial scenarios. But once they are brought into a home living room, the nearly 100,000 - yuan price tag immediately discourages all office workers who want to buy one home - after all, with the same budget, they can hire a real - life nanny for three years, and the nanny can take care of everything.

After all, the tasks they can perform are still too few, and the price is too high. They can't compete with humans in terms of service quality and price, and they are not as convenient and easy to use as large - scale AI models. There may be an inappropriate analogy - when the cost of using slaves is much lower than that of machines, southern plantation owners won't embrace industry.

So, when the cost and revenue don't match, an "Internet bubble - like" loss phenomenon occurs - in 2024, the average gross profit margin of Chinese humanoid robot enterprises was only 18%, and the cumulative loss of UBTECH from 2020 to the first half of 2024 was as high as 4.3 billion yuan.

03

Where Is the Breakthrough?

However, compared with its "predecessors", embodied robots are an "exception".

If you study deeply, you will find that the fundamental reason why those "predecessors" lost popularity is: the concept came first, the technology lagged behind, and the practical value was low.

Take VR glasses as an example. Brands like PICO and iQiyi Qiyu have hit the wall one after another.

The main reason is that the core components are monopolized by a few suppliers. So, there is not only a lack of motivation for development and iteration, but the price also remains high, making them "toys for the rich". When the number of users is limited, the pay - back period for developers is long, there is little high - quality content, and users are "excited before buying but let the product gather dust after buying", finally forming a vicious cycle of "low output - low purchase - low income".

Coupled with the heavy feeling of a headset weighing over 500 grams and the experience flaw of getting hot after half an hour of use, VR is destined to remain in the small circle of hardcore gamers.

Similarly, although the concept of the metaverse is grand, it also cannot bypass the three mountains of computing power, content, and terminals.

So, after the hype, only the concept remains.

However, although embodied intelligence also has the "science - fiction" aura, the mountains in front of it can be moved.

In terms of technology, the rise of embodied intelligence has always been in sync with the iteration of large - scale AI models - the evolution speed of this "smart brain" is driving the embodied intelligence industry forward at an unprecedented pace.

For example, the "Huisi Kaiwu" platform of BAAI tries to solve the problem: It uses a large - scale AI model to drive the "brain" to plan tasks and data to drive the "cerebellum" to control the limbs, and for the first time, it realizes unified control across robotic arms and humanoid robots.

Another example is the cooperation between GAC Group and Huawei Cloud. Through the multi - modal capabilities of Pangu, it can restore the 2D video and 3D point - cloud data of complex driving scenarios within a few minutes, shortening the end - to - end model iteration cycle to "one version every two days".

Are the hardware components too expensive? Enterprises immediately change the plan - for reducers, drivers, and sensors, they switch to the cheaper suppliers, and domestic alternatives can immediately meet the demand.

The Panasonic HG - C laser displacement sensor costs at least 1,500 yuan. The domestic HC6 sensor of Hongchuan Technology can achieve a measurement range of 800 mm and a sampling rate of 2,000 Hz, and the maximum price is only 899 yuan.

To encourage the research of embodied intelligence, Guangming District of Shenzhen directly announced that starting from 2025, it will give a maximum reward of 20 million yuan to newly - settled top 100 software enterprises, top 100 Internet enterprises, and top 100 comprehensive strength artificial intelligence enterprises in the country; and a maximum reward of 10 million yuan to newly - settled "key software enterprises encouraged by the state".

Is the scenario implementation slow? Enterprises are adapting to the market. For example, UBTECH's Walker is used in the education industry, and 60% of its income comes from AI interest classes in primary and secondary schools; Unitree's G1 robot directly won an order of 500 units from BYD in 2024, simply because its price is one - tenth of that of imported products.

On the other hand, major enterprises have started to develop specialized large - scale models for embodied intelligence, such as Helix of Figure AI, which has achieved a cross - scenario task success rate of over 90%.

(Education Robot)

More importantly, China's unique industrial cluster advantage has paved a golden path for embodied intelligence to move from the laboratory to the production line.

In the Guangming Science City of Shenzhen, UBTECH's super factory, DJI's innovation center, and Huawei's HarmonyOS laboratory are clustered, forming a 15 - minute ecosystem of "R & D - testing - mass production". The patent density of robot enterprises is as high as 1,200 per square kilometer, three times that of similar parks in Silicon Valley.

In the Yangtze River Delta, a three - hour drive can cover core suppliers such as Suzhou Green Harmonic's reducers and Ningbo Zhongda Leader's servo motors; Shanghai and Hangzhou's Zhijiang Laboratory jointly established the "Human - Machine Collaboration Joint Laboratory", which processes 100 million pieces of multi - modal interaction data every day. In addition, more than 10,000 robot patents in the Yangtze River Delta are shared for free, and JAKA Robotics has directly saved 30% of its R & D cost, which can be fully invested in the next iteration.

The special feature of embodied intelligence is that it requires both hardware breakthroughs like VR and software revolutions like large - scale models. Its future is not determined by the concept but by the speed of technological iteration - with each upgrade of the large - scale model, the "brainpower" of robots takes a leap; with each replacement of domestic components, the cost drops.

Do you remember how people summarized DJI's success experience before? - DJI was successful because it was located in the Pearl River Delta. Every time it tested and replaced a part, the iteration speed was fast and the trial - and - error cost was extremely low. For any foreign company, when developing a consumer electronic product, just the customization, testing, and selection of components would take three to five times as long as DJI.

When the upstream and downstream of the industrial chain can be independently controlled, China's industrial foundation is the greatest confidence for embodied intelligence, enabling it to quickly transform from a vague concept into a productivity revolution in the physical world.

This article is from the WeChat official account "Shenmou Finance" (ID: chutou0325), author: Gao Teng, published by