StartseiteArtikel

Wenn man nur einen Zeitraum von drei Jahren betrachtet, kann sich die Embodied AI auf keinen Fall rentieren.

36氪的朋友们2025-09-29 12:40
Auf welcher Logik beruht denn die "Rechnung" der Anleger?

In this year's venture capital market, "Embodied AI" has almost been the hottest topic.

According to statistics from CVSource, a division of ChinaVenture Group, there were 138 financing events in the Embodied AI sector in the first eight months of 2025, including 52 financing events worth over 100 million yuan. Just in the week from July 7th to 13th, six Embodied AI startups announced financing of over 100 million yuan. They were Xingdong Jiyuan, which received nearly 500 million yuan; Yunshenchu and Kuawei Intelligence, which received hundreds of millions of yuan; Xiaoyu Manufacturing, which received over 100 million yuan; Xinghaitu, which received over 100 million US dollars; and Tashizhihang, which completed a financing of 122 million US dollars.

In addition to startup financing, in the same week, Zhiyuan Robotics acquired 67% of the shares of Shangwei New Materials, a listed company on the STAR Market, sparking discussions about "backdoor listing" and showing strong potential to become the "first Embodied AI stock." A few days later, Unitree Technology officially started the listing guidance process, making the "first Embodied AI stock" a hot topic.

According to industry conventions, these cases, news, and figures indicate a "cycle" and a "trend."

However, if you dig deeper, the first emotion you might feel is "urgency." Most of the star companies mentioned above completed two or more rounds of financing in the first eight months of this year, making their financing rounds seem rather abstract. For example, the financing round Tashizhihang received in July was positioned as the "Angel + round" - the "Angel round" without the plus sign took place in March 2025, when they set a record for the largest angel - round financing in the Chinese Embodied AI industry with a scale of 120 million US dollars. Xinghaitu completed the "A4 and A5 strategic financing" in the "Series A financing" - this Series A financing started at the beginning of 2025, and the press release stated that the valuation of the latest round had increased by over three times compared to the beginning of the year.

The latest case is Zhipingfang. A press release issued on September 1, 2025, showed that it had finalized seven rounds of financing within half a year, and the scale of each round of financing reached 100 million yuan.

So, the questions arise: Why has Embodied AI become so popular? Who is in a hurry? Where is all this money going? And what kind of logic do investors use to calculate their "accounts"? I think it's necessary to invite investors who are on the front - line and have participated in this wave of phenomenal financing to have a good chat.

Guest for the conversation: Fang Zhenghao, Managing Partner of Xiaomiao Langcheng, who started investing in Embodied AI in 2019

The core of Embodied AI is AI, not machinery

Pu Fan: Your WeChat profile picture is of a Transformer. This makes me recall that "Dabeiyao 14F" first paid attention to Embodied AI in 2023, when the first humanoid robot company, Ubtech, was just established. At that time, we were worried that the term "humanoid robot" was too unfamiliar and would affect traffic, so we deliberately used the title "Investors Can Now Invest in Transformers" to attract attention. Two years have passed, and the situation has completely changed. Humanoid robots no longer need to rely on any hot - topics, which is really thought - provoking.

(Brother Hao's profile picture)

Fang Zhenghao: My profile picture is of Optimus Prime. I changed it in 2022, before Elon Musk named his humanoid robot Optimus. I changed my profile picture because I was very optimistic about the future opportunities in the robot market.

Pu Fan: Let's start from the beginning. You mentioned that you were optimistic about the humanoid robot sector in 2022 and started investing in Embodied AI in 2023. What made you decide that Embodied AI was worth investing in? After all, "liking" and "making a business" are two different things.

Fang Zhenghao: The core reason was the iterative improvement of the artificial intelligence capabilities of GPT - 3.5 and GPT - 4. Firstly, in terms of conversation and human - machine interaction, they can provide a user experience far superior to that of the previous generation of AI - they can have complex, multi - round, and logical conversations with humans. In the previous CNN (Convolutional Neural Network) era, multi - round conversations in NLP (Natural Language Processing) were always a challenge and could only rely on pre - set and pre - arranged content. This means that the interaction ability between robots and humans will be greatly enhanced.

Secondly, thanks to the progress of multi - modal large models and reinforcement learning technologies, the movement ability of robots has been significantly enhanced. At the same time, with the development of multi - modal models, the interaction ability between robots and the physical world is expected to achieve AI generalization. This is fundamentally different from the previous generation of robots. In the past, robots could only complete single tasks in a fixed single - scenario through programming and teaching, but now, the scenario adaptation ability of robots is gradually emerging.

So, from the end of 2022 to the beginning of 2023, the key was to witness the evolution of large - model technologies, which not only optimized the human - machine interaction experience but also broke through the scenario limitations of robots. Artificial intelligence is the core investment direction of our fund, and robots are the ultimate form of AI changing the real physical world - this is the fundamental reason why we started paying attention to the Embodied AI sector in 2023.

Pu Fan: It's interesting. In 2023, the keyword in market discussions was "humanoid robot," and the focus was on topics such as "flexibility," "bipedal or octopedal," and "ability to transform." But you focused on Embodied AI from the very beginning, rather than humanoid robots. So, I'd like to ask further, in your opinion, what is the core difference between humanoid robot companies and Embodied AI companies?

Fang Zhenghao: Firstly, humanoid robots can be defined as "bipedal anthropomorphic robots," with the core features of a bipedal structure and human - like limbs. Among all animals, only humans can support their bodies with four limbs and walk upright on two feet. Even primates cannot do this. The advantage of this physical characteristic is that it can be maximally compatible with human society - the facilities in human society (such as furniture and tools) are designed according to human height and limb movement range, and the specifications have little difference regardless of region or ethnicity.

"Embodied AI" is a borrowed term from the English "Embodied AI." In essence, the core of Embodied AI is still "AI," but it exists in an "embodied" (physical carrier) form, that is, "AI carried on a physical carrier." Therefore, its core driver is artificial intelligence, rather than the "control + remote control" model of traditional robots - AI is the key variable in this sector.

In other words, it is possible to make remote - controlled humanoid robots without AI. Most products of current humanoid robot manufacturers still come with a remote control. Without breakthroughs in AI technology, the uses of such remote - controlled robots are basically limited to demonstrations (such as participating in robot sports games or performances) and cannot solve practical problems. This is also the reason why we invested in Embodied AI from the beginning: we believe that AI is the core of the sector, not the mechanical structure.

The "adaptability" in 2019 had the shadow of "Embodied AI"

Pu Fan: In Xiaomiao Langcheng's investment portfolio, are there more Embodied AI companies founded by large - model teams or robot mechanical teams?

Fang Zhenghao: Teams with an AI background account for a higher proportion. Even in companies involved in ontology manufacturing, the team members include talents in the fields of machinery and control, but the core driver still focuses on AI. The core competitiveness of most teams lies in the AI aspect.

Pu Fan: Which was the first Embodied AI company you invested in?

Fang Zhenghao: Feixi Robotics. We invested in it in 2019.

(The manifestation of force control is using a robotic arm to shave)

Pu Fan: So early? 2019 sounds like "last century."

Fang Zhenghao: Yes. In its early days, Feixi focused on adaptive force - controlled robotic arms, with two core advantages. One is force - control technology, and its team can achieve the most precise force - feedback control in the world. The other is adaptability. The robotic arm can learn and adapt according to different environments on - site, reducing the workload of human teaching and programming. Actually, the concept of "adaptability" from the beginning had the "multi - scenario generalization ability" required by today's Embodied AI. Professor Lu Cewu, the head of Feixi's AI team (currently the deputy dean of the School of Artificial Intelligence at Shanghai Jiao Tong University and the deputy dean of Shanghai Chuangzhi College), is a disciple of Professor Fei - Fei Li from Stanford University. As a co - founder of Feixi, he hoped to create a "robotic arm with generalization ability" from the very beginning, which is consistent with the core logic of today's Embodied AI.

In 2023, we evaluated nearly 100 Embodied AI and humanoid robot teams throughout the year, including companies originally making robot dogs and newly established teams. Finally, we invested in three projects: Qiongche Intelligence, West Lake Robotics, and Songyan Power. Among them, Qiongche Intelligence was founded by Professor Lu Cewu and is in the same lineage as Feixi. Currently, Feixi focuses on adaptive force - controlled industrial robotic arms, while Qiongche Intelligence focuses on general - purpose Embodied AI brains.

Pu Fan: It seems that Feixi follows the logic of industrial robots.

Fang Zhenghao: That's right. Feixi's core capabilities can address scenarios that other industrial robots cannot cover. For example, in processes such as precision assembly (such as inserting computer boards) and grinding, these tasks require the "hand - eye coordination" of skilled workers and the ability to sense the "feeling" during operation - no other robot company has this ability, and Feixi is the first company in the world to cover such scenarios.

At that time, we were optimistic about it for two reasons. One was the AI adaptive learning ability, and the other was the force - control technology. These two factors enabled it to enter the industrial market that traditional industrial robots could not access. At the same time, these scenarios also require "hand - eye - brain coordination," which tests sensor technology and depends on control and AI capabilities. Feixi has advantages in these aspects.

Xiaoye Jiang: Suppose you hire an assembly worker in Vietnam, and the monthly salary may only be a few thousand yuan. Using Embodied AI to solve such problems, won't it seem "too costly"?

Pu Fan: I also want to add a question. In the actual application of Feixi robots in factories, is there any data to prove the improvement in production efficiency? What is the improvement rate? Has its revenue scale entered the right track?

Fang Zhenghao: Objectively speaking, currently, robots' cycle efficiency (the amount of work completed per unit time) in handling complex tasks in complex environments is still lower than that of skilled workers. However, you should know that skilled workers need to work in three shifts, while robots do not need to rest, drink, or eat, and they will not have emotional problems or make mistakes due to long working hours. The core advantages of robots are "stability" and "continuous operation ability." So, even if the single - cycle efficiency is slightly lower, in most scenarios, the ROI of robots has exceeded that of humans.

The issue of "cost" should be viewed from a long - term perspective. In the past, the core advantage of China's manufacturing industry was the "demographic dividend + engineer dividend," especially in the low - and middle - end manufacturing industries, which highly relied on the demographic dividend. However, now, the demographic dividend is gradually disappearing. From many data, we can judge that in the next decade, there will be a huge labor shortage in many domestic fields. The rise in labor costs is an irreversible trend. This situation has already occurred in Europe and the United States, which is also the fundamental reason why they attach importance to the robot industry, aiming to solve the problem of manufacturing hollowing - out.

Relocating factories to Vietnam, Africa, or South America is indeed a phased solution. At the same time, the capabilities of robots (robustness, reliability, and cycle efficiency) are continuously improving, and the cost is constantly decreasing. Under the trend of large - scale manufacturing, the gap in the "investment - output ratio" between robots and humans will gradually narrow and even reverse, which is also irreversible.

In addition, Europe and the United States are promoting the return of manufacturing. If China wants to maintain its manufacturing advantage, it must establish advantages in the "technological capabilities" and "industrial chain discourse power" of the robot industry, which is an inevitable long - term choice.

Pu Fan: Moreover, China has complete supporting facilities and an abundant engineer dividend. A strong supply chain can support the stable operation of robots, and a rich variety of industrial scenarios can enable robots to be quickly put into use. Maybe the cost advantage will soon become apparent?

Xiaoye Jiang: In my experience, it may still take a long time. I once visited a pharmaceutical company, and there were many robotic arms operating in its R & D base. However, the temperature on - site was very high, and the robotic arms still needed manual verification. In the future, it may be a model of "robotic - arm - dominated operation + engineer - team monitoring," similar to the current "assembly - line + manual inspection" model in many factories.

Fang Zhenghao: The scenario you mentioned is actually "automation in standardized scenarios." If the incoming materials in manufacturing are stable and the processes are standardized (whether it is process - type manufacturing or discrete - type manufacturing), automation can actually be achieved long ago, without the need for Embodied AI or even robots. An automated production line can meet the demand. For example, in scenarios such as food and beverage packaging, pre - cooked food boxing, and plastic sealing, automation has been popular for many years.

The core problem in scenarios that have not yet achieved automation is the "uncertainty of incoming materials." For example, in express sorting, material sorting, and loading and unloading in manufacturing, the shape and position of materials are not fixed, and traditional automation means and robots cannot fully cover these scenarios. These "non - standard workstations" are the incremental market that Embodied AI can bring, and most of these workstations have the opportunity to be replaced by robots in the future.

Zhu Xiaohu is a typical "pragmatist"

Pu Fan: Let's talk about Songyan Power. It should be the most talked - about Embodied AI company among your invested companies. The "short robot in the Yizhuang Marathon" attracted a lot of attention, and its valuation increased rapidly. I read an interview with its founder. At the beginning of the year, its valuation was only 500 million yuan, and it rose to 2 billion yuan after participating in the robot marathon. Even so, investors still flocked to it.

Songyan Power also has a very eye - catching label. At the beginning of the year, Zhu Xiaohu said that he was "withdrawing from the humanoid robot field in batches," and Songyan Power was one of them.