Embodied Intelligence: Precipitate in Fervor, Break Through in Implementation

In the second half of the artificial intelligence era, embodied intelligence has become a hotly contested territory. At the Consumer Electronics Show (CES), a major tech event, robots can be seen everywhere. There are vacuum cleaning robots, lawn mowing robots, agricultural robots, and countless humanoid robots...

On one hand, capital and industry giants are collectively placing bets, and leading companies under the spotlight are raking in profits in batches. On the other hand, small and medium - sized enterprises are facing a financing gap, and some star companies are falling into crises. It's a situation where calmness and cruelty coexist.

Undoubtedly, after a short but intense battle, this race is shifting from "competing for financing and showing off technology" to the deeper waters of "comparing real - world applications and profitability."

The Contrasting Scenarios after the Influx of Hot Capital

In the embodied intelligence track in 2025, the capital market presents a unique picture of "half seawater, half fire."

First of all, in terms of industry popularity and financing scale, the macro - fundamentals of the track are undoubtedly hot.

According to IT Juzi data, as of December 21, 2025, more than 600 investors have "genuinely" increased their investment in embodied intelligence. There have been over 304 financing events throughout the year, and the total financing amount has reached 37.9 billion yuan, which are 2.95 times and 4.05 times that of 2024 respectively, and 7.24 times and 4.37 times that of 2023. Among them, there is even the spectacular scene of Songyan Power completing 6 rounds of financing within a year, rapidly advancing from Series A to Pre - B+ round.

Large - scale companies and industry giants have also collectively entered the fray, making high - frequency and focused investments, and have been extremely prominent for a while.

According to incomplete statistics, 8 core large - scale companies such as Baidu Ventures, Lenovo Capital and Incubator Group, and Ant Group have made a total of 62 investments throughout the year. Among them, Baidu Ventures leads with 13 investments, followed closely by Lenovo Capital and Incubator Group/Lenovo Star with 11 investments. Guoxiang Capital (SenseTime) and Ant Group are tied for third place with 8 investments each, forming the first investment echelon.

In the wave, Tencent led the investment in Zhiyuan, which was also its first foray into the humanoid robot field. It quickly increased its investment after a month and soon became a shareholder of Unitree. This shows its strong focus on embodied intelligence and determination to enter the game.

It's worth mentioning that Wang Xing of Meituan has made concentrated investments through its two investment platforms, Strategic Investment and Longzhu. Since 2023, it has made a total of 8 investments, covering 7 core companies. In the first half of 2025, it made 6 investments in a concentrated manner, investing in 2 companies within 11 days before July. It appeared 9 times in the leading echelon, with a frequency and concentration ranking among the top in the industry, firmly securing the position of "China's No.1 investor in embodied intelligence."

JD.com is also unstoppable. It has rapidly and deeply deployed by investing in 6 robot companies within three months! Different from Tencent's core logic and investment boundary of "model + hardware" dual - wheel drive (Tairos large - model empowering Zhiyuan humanoid robots and Unitree quadruped robots), JD.com adopts the approach of "scenario + full - chain." Its six investments cover the robot body (Zhiyuan, Pasini), core components (Zhujidongli), and algorithms (Qianxun Intelligence, Zhongqing Robot, RoboScience), directly serving cost - reduction and efficiency - improvement in logistics processes such as sorting, handling, and last - mile delivery. According to industry analysts' estimates, the total scale of JD.com's six investments is nearly 4 billion yuan. It can truly be said to be making a "powerful move"!

The logic behind the frenzy of large - scale companies and industry giants is simple: after the AI wave, embodied intelligence is regarded as a terminal form that may change the world, and no one wants to miss out on "the next Unitree."

However, behind the frenzy, there are also different "figures."

Let's first look at a set of statistical data from the Touzhong Jiachuan team: among the 168 embodied intelligence companies that received investments in 2025, the top 10 companies, including Zibianliang, Tashizhihang, Leju Robot, Xingdong Jiyuan, Xinghaitu, and Zhongqing Robot, received a total of 13.472 billion yuan in investments, accounting for 40.95% of the annual industry financing amount.

The "money - sucking" ability of leading companies has led to an increasingly obvious trend of capital concentration. The result is obvious. The narrowing of the financing window for mid - tier and lower - tier companies also indicates that the industry structure may change, and the acceleration of stratification is only a matter of time. In fact, this stratification has already become apparent. The "exit" events after struggling are an inevitable result of the survival - of - the - fittest rule in the industry. Small and medium - sized enterprises are facing a financing gap, and some companies have not received new funds for 2 - 3 years, struggling under the dual pressure of technological R & D and mass - production investment.

Take Yunji Technology, which was once very popular. At that time, it ranked first in the Chinese robot service intelligent agent market with a 6.3% domestic revenue share, and its "Run" series of robots became a standard in hotels. However, it has not received new financing since Series D, and the financing gap has exceeded 3 years.

The unicorn humanoid robot company, CloudMinds Robotics, is also deeply mired in a crisis. This once "unicorn" that had accumulated over 5.4 billion yuan in financing has not received new investment since 2023. The industrial and commercial information on Tianyancha shows that 61 "enforcement cases," 101 "judicial cases," and several "restrictions on high - end consumption" already illustrate the severity of the crisis and the company's troubled future.

There's no need to doubt the sensitivity of capital. Being decisive is its "standard feature."

Zhu Xiaohu, a partner at Jinshajiang Venture Capital, caused a stir when he withdrew from multiple humanoid robot companies in batches. As an investor who once bet on Feixi and Xinghaitu, he withdrew from multiple embodied projects around the beginning of 2025, bluntly stating that "the business model is not clear." This is also a signal that capital is returning to rationality. When the hot money becomes more rational and cools down, the calm thinking will point to "companies without technological barriers and real - world application capabilities," and they will eventually be eliminated by the market.

In contrast, the advantages of leading companies will continue to expand.

In 2025, Unitree received nearly 1.2 billion yuan in orders, Zhiyuan's sales revenue exceeded 1 billion yuan, and Galaxy Universal's orders exceeded 700 million yuan. These tangible achievements have become the "hard currency" for financing. On the other hand, companies that only stay at the conceptual stage and lack real - world application scenarios are gradually being abandoned by capital.

A senior investor admitted, "Now, capital doesn't just look at the technical parameters on the PPT. It wants to see if the robot can operate stably in the factory and make money." Just as many investment managers can't help but ask, "The technical parameters are beautiful, but where are the customers?"

The Crucial Leap from "Showing off Technology" to "Making Money"

"Where are the customers?" Essentially, it's a problem of real - world application scenarios.

Showing off technology at the primary stage of technical theory is not only a ticket to enter the game but also lays a better foundation for upgrading. After all, the model needs to be iterated, training data needs to be accumulated, and control strategies and algorithms need to be optimized. In the process of iteration and upgrading, continuous exploration is being made on the issue of scenario adaptation for "doing work." The anchoring of scenarios also determines the differences in technological R & D, product form, and market strategy.

In the video of the humanoid robot released by Hangzhou Deep Robotics, the robot shows off various skills such as stepping, side - kicking, and doing Tai Chi, amazing millions of netizens. However, in the production workshop of CATL, the "Xiaomo" robot of Qianxun Intelligence is repeating a boring but crucial task - autonomously detecting the connection status of wiring harnesses and dynamically adjusting the plug - in force to avoid damage to high - voltage components. The implementation of this robot can save the factory hundreds of thousands of yuan in labor costs and losses every year.

The comparison between these two scenarios reveals the core logic of the commercialization of embodied intelligence: the leap from "showing off technology" to "practical money - making."

As Mr. Wang Xiang, the former president of Xiaomi and a partner at Gaoshan Xinyu, said, "The advancement of technology doesn't equal commercial success. Finding real needs and application scenarios is the key." Fancy moves can't solve the problem. "Competing to see whose robot can perform more complex actions" is only the primary stage. "Doing work" can create greater value.

Whether it's the more popular humanoid robot track or the wheeled and quadruped robots, players in the field of embodied intelligence are indeed exploring scenarios and conducting industrial practices in a "diverse" manner as mentioned above.

The ultimate goal of humanoid robots is to adapt to general environments like humans. However, currently, they have high costs and face significant technical challenges. Therefore, they are first piloted in structured industrial environments such as automobile manufacturing. For example, the Zhiyuan Expedition A1 humanoid robot is conducting on - site tests and verifications for tasks such as assembly, handling, and material transfer in BYD's factory. This is also a crucial step for Zhiyuan robots to move from the laboratory to the industrial field. It is reported that Zhiyuan has made batch deliveries in scenarios such as automobile parts and consumer electronics assembly, with core customers including Joyson Electronics, Longcheer Technology, and Chery's supply chain. The cumulative order scale has exceeded 100 million yuan.

Wheeled robots have the highest efficiency and stability in flat and regular environments and have been widely commercialized in scenarios such as cleaning and logistics. For example, Gao Xian Robotics, the global leader in commercial cleaning robots, has a series of commercial cleaning robots that can automatically scrub the floor, sweep dust, and disinfect. They have been widely used in shopping malls, factories, airports, etc. in more than 40 countries and regions around the world, with a running mileage of over 300 million kilometers and rich scenario data accumulation.

Real - world application of scenarios has become a real core competitiveness. It's like a final exam, testing the work results of the "school" and the learning effects of the "students." This crucial leap in scenario application not only brings real income but also builds a positive cycle of "data - technology - product." It is also the inevitable path for embodied intelligence to enter the deeper waters.

However, large - scale scenario application as people envision definitely won't happen overnight. From a technical perspective of scenario application, there are still many problems to be solved. We need to give the industry more time and space.

For example, humanoid robots have insufficient adaptability in the unstructured physical world. In shopping mall crowds and complex industrial working conditions, the mis - touch rate of robots still reaches 8%. This is a publicly - reported statistical value from on - site tests by multiple leading companies, which clearly exposes the gap between the "ideal environment" in the laboratory and the "complex variables" in the real - world scenario. As mentioned in a Skynews report about CES, "When further investigating this clothes - folding robot, I found that it needed four days of remote operation to adapt to the new tables and lights in the unfamiliar CES environment."

In addition, the "uncanny valley effect" also restricts the popularization of humanoid robots in family scenarios. This stems from the need to improve users' trust in humanoid robots. This effect is magnified in family scenarios. Coupled with multiple concerns about safety, privacy, and trust, it directly reduces users' acceptance and trust in humanoid robots, becoming a key bottleneck for family - wide popularization. Zhao Weichen, the vice - president of UBTECH for accelerated evolution, pointed out that if a robot's facial features are too realistic but can't fully achieve natural human - like expressions, it may cause discomfort and a sense of alienation among users, creating the "uncanny valley effect."

Moreover, in real - world scenarios, the data from the visual and tactile sensors of intelligent robots is noisy, and there are delays and errors in physical movements, resulting in insufficient mapping accuracy between "instructions - actions." The coordination between force control and dexterous hands is not precise enough. For example, when a robot tries to pick up a cup, it often drops it, which shows that this type of technology is not yet fully mature. For instance, in the video of Tesla's Optimus humanoid robot selling popcorn circulated on social platforms, we can also see the robot "immediately releasing and readjusting when the grip is too strong."

Real - world application of scenarios tests the integrity and practicality of the technical migration of embodied intelligence from the "laboratory" to the physical world. Although intelligent robots perform well in the laboratory environment (structured and interference - free), their performance will decline in open environments (such as homes and outdoors) when facing unknown obstacles and complex lighting. This weak generalization ability is just one aspect of the technical complexity of scenario application.

True large - scale scenario application requires the entire industry to promote it. The first and foremost thing is cost reduction, which is not the responsibility of one or a few companies. It requires ecological collaboration and overall cooperation.

Breaking through the "Last Mile" Is Never a Solo Act

The single - unit price of intelligent robots ranges from hundreds of thousands of yuan to millions of dollars, which poses certain obstacles to large - scale deployment in industrial scenarios and universal popularization in family scenarios.

Taking Tesla's Optimus as an example, its single - unit price is about 200,000 - 250,000 US dollars, and the cost of core components such as sensors, motors, and lead screws accounts for over 70%. This high proportion stems from the extreme requirements of humanoid robots for precise execution and perception capabilities. Compared with Optimus, Boston Dynamics' Atlas, as a research and development prototype, has a proportion of special materials and customized components exceeding 60%. The single - unit cost is as high as 1.5 million US dollars, only meeting the customized needs of military and high - end industries.

Domestic enterprises have significantly reduced some costs through domestic substitution, but the battle to break through the cost - performance bottleneck continues. The underlying reason for the high cost is essentially the lack of industry standards and insufficient collaboration, which is also an inevitable stage in development and needs to be gradually improved.

Industry analyst Zhang Wen believes that there is a lack of unified specifications in interfaces, communication, data formats, and performance evaluation. Enterprises have to fight on their own, and the proportion of general - purpose components is less than 30%, failing to release the scale effect. Moreover, the motors, reducers, and sensors of a single robot often come from different supply chains.

It's not hard to see that from the laboratory to large - scale industrial application, the "last mile" of the implementation of embodied intelligence is mainly stuck at two bottlenecks: high cost and fragmented standards.

To solve the cost dilemma, the key lies in breaking the current situation of fragmented industry standards and driving supply - chain collaboration and component generalization through standardization, Zhang Wen finally expressed his view to WANDIAN RESEARCH.

Standardization is the core means to achieve cost reduction and promote the real large - scale application of scenarios, which is also a consensus in the industry. According to calculations, for every 10% increase in the generalization rate of components, the single - unit cost of robots can be reduced by 8% - 10%, and the adaptation cycle can be shortened by over 40%.

On December 26, 2025, the Standardization Technical Committee for Humanoid Robots and Embodied Intelligence of the Ministry of Industry and Information Technology (hereinafter referred to as the Standardization Committee) was established in Beijing. The current situation of "the increasingly prominent problem of lagging standards, the high collaboration cost caused by non - unified basic interfaces, and the impact on market trust due to the lack of application specifications" will also be gradually improved. At present, standardization work is accelerating under policy guidance. The Standardization Committee has initiated 12 core standards in areas such as environmental perception, operation, and safety specifications, focusing on three key directions: unified interfaces, data interconnection, and performance evaluation. It plans to release the first batch of specifications in 2026, providing policy support and platform guarantee for cost reduction in the industry.

Moreover, the industry is moving from "fighting alone" to "standard - based collaboration."

In the field of component generalization, Unitree Technology has jointly formulated the servo - motor interface standard with core component manufacturers such as Green Harmonic and AMC Electric, reducing the procurement cost of general - purpose components by 35% and shortening the delivery cycle by 40%. Relying on this collaborative achievement, the single - unit cost of Unitree's G1 robot has rapidly decreased after adopting standardized joint modules. The subsequent R1 educational and research model is priced at 39,900 yuan, becoming a standard in university laboratories. Wang Xingxing