Is factory work the only way for robot commercialization?

Great Potential in "Entering Factories"

01 When Robot Companies Start to Go Bankrupt in Batches

In an interview in the first half of the year, Zhu Xiaohu said, "Embodied intelligence is very popular now. At the same time, I think the commercial path is still unclear, especially for humanoid robots. I often joke that now every humanoid robot can do somersaults, but where is the commercialization?" Zhu Xiaohu also said, "We are withdrawing from humanoid robot companies in batches" and "There is a consensus on robots, but the commercialization is unclear."

Zhu Xiaohu's words are like pouring cold water and have sparked fierce debates in the industry. However, behind the hot market, most robot enterprises are in a difficult situation. A slight mistake may lead to the risk of bankruptcy.

K-Bot developed by K-Scale Labs

Since the beginning of this year, the Silicon Valley star company K-Scale Labs has burned through $4 million in financing. The founder sent an email to customers to refund the money and close the accounts. The American children's companion robot company Embodied has directly stopped its product services, and all its employees are unemployed.

The French company Aldebaran, known as the global pioneer in robots, has been working on humanoid robots for more than a decade but finally failed. The supply chain bottleneck has led to high costs. The products are expensive and not easy to use. After losing $29 million last year, it went bankrupt and was finally acquired by a Chinese company to survive.

A recent report from Goldman Sachs is even more pointed. After in-depth research on 9 core Chinese robot suppliers, the research team found that these enterprises are frantically deploying production lines at home and abroad based on the expectation of the industry's explosion. Behind this "capacity first" strategy is the gambler's logic familiar to the Chinese manufacturing industry - betting heavily on the future at all costs to lock in the supply chain dominance with scale advantages.

Goldman Sachs believes that Chinese robot supply chain enterprises have planned a huge production capacity of 100,000 to 1 million units per year, but no company has received large-scale real orders.

So, is Goldman Sachs' report true? Is there any exaggeration? It depends on how to define large-scale real orders.

As of the first half of 2025, from the publicly available orders for humanoid robots in China, the procurement volume of educational and scientific research institutions such as universities, research institutes, and vocational colleges accounts for 75% of the total order volume. The enterprise orders are generally between several thousand and several hundred units, and most of them do not need to be delivered this year. Regarding the specific implementation scenarios, efficiency improvement, and PMF verification, there are basically no positive responses. Therefore, some industry insiders believe that prematurely releasing the signal of "large-scale orders" is no different from triggering a bubble before the industry has taken shape.

The most typical case is that in the first half of the year, Tesla signed a letter of intent (LOI) with the American pharmaceutical company PharmAGRI, planning to deploy up to 10,000 Optimus V3 humanoid robots in the operation of its self-owned farms, the synthesis of active pharmaceutical ingredients (API), and the production of prescription drugs. Apart from Musk's direct answer that it was false, this case is also worth looking into.

First of all, there is no sign that Optimus V3 can be mass-produced on a large scale. Suddenly, a large order of tens of thousands of units appears, which makes people hard not to doubt its authenticity. Secondly, in terms of funds, the scale and financial ability of PharmAGRI do not match this scale of procurement. After all, this means an expenditure of at least hundreds of millions of dollars.

This "ten-thousand-unit order" thrown out by PharmAGRI is essentially a letter of intent (LOI), not a formal contract. From a legal perspective, it does not have mandatory binding force, and it is impossible to clarify the core elements such as the procurement amount, delivery cycle, and payment terms. Such "vague orders" are not uncommon in recent industry promotions. It is quite thought-provoking that PharmAGRI neither responded nor questioned.

Industry experts point out that this is bringing the bad atmosphere of the mobile phone and automobile industries to the embodied intelligence industry. The embodied intelligence industry is still in an obvious early stage. Excessive marketing and publicity, inflating the valuation, will only damage the trust foundation of users and the industry. The real delivery situation of enterprises should be verified by comparing orders with actual delivery quantities, tax payment data, etc., and enterprises with serious data deviations should be punished.

02 The Era of Robots Entering Factories Has Arrived

But is there really no way for embodied robots to achieve commercialization? Compared with entering the home scene to perform complex and diverse tasks, "entering factories" is really one of the best choices at present.

For small and medium-sized robot enterprises, the only way left for them is to enter factories and compete in the industrial scene implementation. Some enterprises in the industry signed many project contracts in the early stage of development, including some projects with low profit margins and relatively strict requirements. Later, they found that they could not digest these projects, and the company encountered difficulties in product iteration and financing. Therefore, for robot industry chain enterprises, it is not about how high the valuation is, but about how long it takes to become the first in the segmented market.

Now, most enterprises that have reached the Series B and later stages have two characteristics: first, they have clear commercialization scenarios (for example, Unitree focuses on logistics and warehousing, and Galaxy Universal specializes in industrial inspection), second, they have achieved small-scale delivery (annual revenue exceeds 50 million yuan). For example, the revenue of Zhipu Robotics reached 120 million yuan in 2024, making it one of the few embodied intelligence enterprises with definite revenue.

So, what difficulties need to be overcome to achieve these two goals?

To adapt to multiple tasks, general-purpose robots must retain a large number of redundant degrees of freedom, sensors, and computing resources, which is a waste of cost in the industry. Moreover, customers prefer "specialized machines for specific purposes" and optimize the minimum closed-loop for specific tasks rather than paying for general capabilities.

First of all, it is necessary to customize models and hardware for customers. The industrial scene data is relatively closed and difficult to share. Even if you persuade the customer, there is still the problem of customized production of robots, optimizing the minimum closed-loop for specific tasks rather than paying for general capabilities. Isn't it just customized optimization? Who among those who have worked as a service provider doesn't understand? But you may be stunned when you really do it. Take the automobile assembly line as an example. The robotic arm needs to tighten bolts with millimeter-level precision (±0.3mm) and have a microsecond-level response speed. The greater challenge lies in the adaptation to the dynamic environment: when the conveyor belt speed fluctuates or the workpiece position shifts, the robot must adjust its actions in real time. Tests by a leading enterprise show that the success rate of its latest model is as high as 98% in a static environment, but it drops sharply to 67% after introducing dynamic interference, highlighting the technical bottleneck in practical applications.

Secondly, the industrial scene requires the stability of robots to reach six sigma (6 Sigma), which corresponds to a pass rate of 99.99966%, meaning that in every million opportunities, the defect rate does not exceed 3.4 times, almost reaching the level of zero defects. From the current technical path of general-purpose humanoid robots, the success rate of grasping actions trained by generative AI is between 80% - 90%. This is absolutely unacceptable for the industrial scene, which means that there may be tens of millions of losses every day.

Regarding this, Yikeli Information Technology, which is deeply involved in the embodied intelligence business, shared at an industry internal salon that currently, robots in the industrial scene can only perform fixed tasks and cannot effectively and efficiently perform dynamic tasks. Moreover, embodied robots have clear reference objects, such as industrial robots and collaborative robots. The hardware load capacity of embodied robots is only in the primary stage, and its speed is 5 times slower than that of industrial robots. In addition, the accuracy of the code generated by the current large model cannot meet the industrial requirements, and it cannot handle sudden situations. Without network connection, the model accuracy will further decline.

Industry authorities believe that when a company's annual shipment volume exceeds 50,000 units, it not only means a decrease in the cost space but also means a huge change in the industry's recognition of the solution. At present, only logistics robots are relatively close to this goal. There is still no sign of a turning point for general-purpose robots. Perhaps in the industrial segmented scenes, there will be robot companies that are efficient enough and have low costs to stand out.

Conclusion

Although the warning in the Goldman Sachs report is a bit exaggerated, it also reveals that in the short term, embodied robots are not yet ready to replace humans. The orders from leading manufacturing enterprises are more about "strategic layout for the future + trial and error on the basis of automation" and the competition for future discourse power, rather than rational transactions based on a proven business model.

There is no need to be overly alarmed by the mention of bubbles. From the perspective of the long economic cycle, historical experience has told us that in the wave of technological revolution, moderate bubbles are exactly the special catalyst to promote the progress of civilization.

If you are a robot or AI practitioner or interested in them, we can invite you to a high-quality communication group. Please note your personal information when applying to join.

Reference materials:

Thousands of Million-Level Orders for Humanoid Robots Are Piling Up. How Much Bubble Is There? Source: Humanoid Robot Release

Who Is the Financial Backer Behind the Trillion-Dollar Robot Track? Source: Caijing Magazine

Transform into the First Humanoid Robot Stock in 120 Days. Source: Tencent Finance

This article is from the WeChat official account "Chief Innovation Review Bureau", author: Zuojing Guantian. Republished by 36Kr with permission.