Most embodied intelligence entrepreneurs don't really understand PMF.

Recently, I've heard two of the most talked - about things in the embodied intelligence circle:

Firstly, as long as a product is launched, someone will buy it. So what if its lifespan is only 100 hours?

Secondly, Company Y in the field of embodied intelligence, which had originally reached a cooperation with the international auto parts giant Company B and was valued at over 10 billion RMB, was kicked out of the factory due to unmet delivery expectations.

I can't help but sigh that the current embodied intelligence track is as bustling as Chinese New Year. Hot money is flowing, and concepts are flying everywhere. Many companies are extremely busy, and investment institutions looking for business plans on the streets are a force not to be underestimated. And for entrepreneurs, the words that come out of their mouths are either "Scaling Law" or "General Artificial Intelligence".

However, perhaps only those deeply involved in the industry truly understand the real situation.

Although the wave of embodied intelligence has been blowing for a long time, an investor told me that the vast majority of entrepreneurs in embodied intelligence still haven't found the way to Product - Market Fit (PMF). After all, when it comes to actually taking the product out of the laboratory and putting it on the shelf for sale, many of them will be at a loss.

The so - called PMF is the abbreviation of Product Market Fit, which means that the product and the market reach the best fit point. A company's product exactly meets the market demand and satisfies the customers. This is the first step towards a successful startup.

Although PMF has different meanings in different fields, in my humble opinion, in the robot market, at least some real users need to use the product for a period of time to know whether the product fits or not. And this "fit" is neither equal to orders nor equal to curiosity.

Although humanoid robots are everywhere now and the financing news is getting louder and louder, many companies are actually using the old map of autonomous driving to find the new continent of embodied intelligence. This is also a well - known practice in the industry.

For example, many companies are still struggling with how to make robots walk more gracefully and anthropomorphically, thinking that the more data they pile up and the larger the model, the more general it will be. However, they ignore the ability of robots to have the common sense of life and physical intuition that they should have in the three - dimensional world.

Pricing

In the business closed - loop, pricing is the best litmus test. It is not only a number but also represents a company's underlying understanding of the product's value.

Recently, I heard a joke. A team priced a consumer - oriented device at $15,000 and showed it to Silicon Valley VCs. The partners were not convinced and were even reluctant to spend the money that is usually willing to be spent on trying new things.

Does this mean that the current embodied intelligence market has been torn into two extreme price ranges?

This is just the tip of the iceberg of the current strange situation in embodied intelligence. On one hand, there is emotional consumption ranging from a few hundred to a few thousand yuan. Compared with functions, users are more buying happiness and companionship, like those small desktop robots or programming teaching aids. Even if they buy the wrong one, they won't go bankrupt, and the tolerance for trying new things is relatively high. On the other hand, there are productivity tools costing tens of thousands or even hundreds of thousands of yuan. In this price range, bosses are buying not high - tech but Return on Investment (ROI). As long as the robot can screw bolts continuously for 24 hours and replace two shift workers, and the cost can be recovered in a year, then the price of hundreds of thousands of yuan is really worth it.

The most embarrassing are the products priced at around $10,000, equivalent to about 100,000 RMB. This middle - range price is neither low enough to attract mass consumers nor high enough to replace labor and create a clear ROI.

Theoretically, buying a top - level drone for $10,000 - 20,000 can take pictures from a god's - eye view and can be regarded as social currency. But if you buy a robot for over $10,000 and it can't even carry a cup of tea steadily, then in fact, it is just a beautiful but useless industrial waste taking up space in the living room.

For this reason, an investor sighed that many entrepreneurs simply add a profit margin to the hardware cost and push the product to the market, thinking that this is the fair market price. However, this has nothing to do with PMF. It is just what cost accountants should do.

Orders

Speaking of those so - called glamorous orders, most of them are actually about social relationships.

Many star startups, in order to make the revenue curve in their business plans look better and get a higher valuation in the next round of financing, have an unspoken rule: sell robots to their upstream suppliers.

Although this sounds quite black - humorous, it has been recognized as a sustainable business model of resource exchange. For example, in order to maintain a relationship with these future potential unicorns or to cooperate with the investors in their PR activities, suppliers will symbolically buy a few robots and put them in the exhibition hall.

As a result, the money in the left pocket is successfully transferred to the right pocket.

However, this is not a demand derived from real operations in factories and warehouses. Therefore, this model is doomed to be unsustainable.

In my years of writing career, after contacting many manufacturing bosses, I found that they are actually the most realistic group of people because they only care about ROI. "If I spend 200,000 yuan to buy a humanoid robot, and its work efficiency and stability are not as good as a non - standard automated bracket or a robotic arm that costs 20,000 yuan, why should I buy it?"

The current situation is that many scenarios actually lack a dexterous hand that can accurately grasp objects and even peel eggs. However, in order to raise a large amount of money and tell a big story, entrepreneurs prefer to sell the seemingly high - end but actually redundant and bulky robot body.

This obsession with the humanoid form actually shows that they haven't seen through the essence of industrial efficiency.

You may wonder why these extremely smart entrepreneurs know that the humanoid shell is bulky and the PMF is unclear, but still rush into it one after another?

As it is said in "Records of the Grand Historian": Everyone in the world knows that something cannot be done, but no one dares to say it. Due to the lack of long - term product definition ability, people keep their consensus in their hearts and can only rely on technological speculation.

Defining a niche tool with real ROI is extremely difficult. It requires you to have real respect for factories, workshops, and the complex physical world. In contrast, putting on a humanoid shell and telling a grand story about a general artificial intelligence carrier is the easiest shortcut to get investors to take out their money.

Because they can't see the real target clearly, they simply make the situation grander, trying to cover up their strategic laziness with the illusion of scale.

Luck

Regarding speculation, I've heard a metaphor from other seniors in the past two years: Most embodied intelligence companies are actually playing the "pistol mode".

What does it mean? It means they only have a few bullets in hand, take one step and look around, aim by feeling and luck, and even don't know where the target is.

You can judge whether a company is relying on luck like this: look at its R & D plan. If it thinks it can achieve success in one go without leaving any room for failure and iteration, then it is basically shooting blindly. These companies generally like to say that this is rapid iteration, but behind closed doors, everyone knows that in fact, they are gambling, hoping that a certain iteration will lead to an epiphany or a certain scenario will work out.

However, this is not a long - term solution. It is actually no different from gambling.

If it is a large company, it can afford to suffer a few losses by treating data as the key factor and desperately piling up data. After all, it has a large capital reserve and can afford to invest in computing power and manpower. It can look for the so - called emergence in the blind boxes that sometimes work and sometimes don't. At worst, it can start over and then say externally that it will replace the laboratory director with a person with a better background and start again.

But it is not easy for a startup team to do this. If they follow this approach, they may burn out their cash flow easily.

Especially those leaders with a strong academic background often overly believe in the theoretical beauty. The large - model teams are always staring at the Scaling Law, thinking that piling up computing power and data can create miracles. The hardware teams are sticking to the old - school PID control theory, researching how to make the motors more refined.

The most ridiculous thing is that these two groups of people can't even have a good conversation. The final product is like a Frankenstein, with strong and expensive "muscles" but a slow - reacting and under - developed "brain".

It's really absurd to expect such a product to work in a complex real - world environment.

Inertia

Many of these embodied intelligence startup teams bear the mark of autonomous driving or leading consumer electronics companies.

Of course, this is both an advantage and a trap. These people do have strong engineering capabilities, but in the robot field, the "hammer" in their hands is too heavy, making them see everything as a "nail".

The basic logic of autonomous driving is to achieve the closed - loop iteration of algorithms through a large amount of road - test data on two - dimensional (or quasi - three - dimensional) structured roads.

However, robots face an unstructured and highly interactive three - dimensional environment. Even walking a few steps at home, the complexity of the obstacles they encounter is several orders of magnitude higher than that on a highway. If a car crashes, there is insurance, but if a robot falls or breaks, it may mean the stagnation of the entire R & D cycle.

Many entrepreneurs always want to replicate the data flywheel of autonomous driving, but the reality is that before the flywheel starts to turn, the hardware purchased at a high cost may fall apart first.

Some well - known large companies are extremely good at focusing on details, even being very particular about the damping of each joint and the routing of each cable. However, this extreme craftsmanship fails to work when a company is transforming towards an AI strategy.

The root cause is that they always use a hardware - oriented thinking to do things and then let the algorithm adapt. But in the era of embodied intelligence, this logic obviously doesn't work. It should be that the algorithm, which is like the brain, is sorted out first, and then the most streamlined and suitable "body" is configured.

"You should first let the algorithm train in a virtual environment, figure out what kind of perception and feedback it needs, and then customize the most streamlined and obedient body in reverse. If you still cling to polishing the robot body, it's like forcing a heavy armor on an unformed soul." said a private enterprise owner.

"If you still stick to the old - fashioned inertial thinking, you will really be looking for a new continent with an old map."

Finally: Some immature observations

Ultimately, the current bottleneck of embodied intelligence does not lie in whether the computing power is sufficient or whether the algorithm is effective, but in the lack of understanding of industry division of labor.

When you can't figure it out, just look at history. The current situation is a bit like the early days of the microcomputer era: every company has to self - develop joints, motors, reducers, and sensors. This highly vertical and integrated path seems to control everything, but most of the startup team's energy is consumed in making parts, and they have no time to figure out what the users really want.

Robots that can do backflips and side - flips at press conferences are really eye - catching, but they are worthless in the real business environment. Commercial customers never care about whether it can do another flip. Instead, they care about: whether it can work stably, whether it can replace a certain amount of human labor, and whether it can reduce the marginal cost low enough.

An authoritative report clearly states:

In 2025, the global commercial service robot market size is about $8 billion. It is expected to grow to about $30.8 billion by 2034, with an annual compound growth rate of about 21.3%. High deployment costs and adaptability issues are important obstacles to rapid adoption. Many small and medium - sized enterprises find it difficult to prove the economic viability (ROI). Due to the high threshold, many small and medium - sized enterprises are slow to deploy. The market growth comes more from the actual value of efficiency improvement and labor substitution, rather than the technological show - off itself.

Even after all these years of development in autonomous driving, the players still in the game are not the companies that are best at making demos. Similarly, in the robot industry, the enterprises that understand social division of labor and the marginal cost curve will be the first to succeed.

Because they can be the first to get out of the engineering quagmire of making everything by themselves, hand over the non - differentiated parts to the mature supply chain, and concentrate limited resources on the capabilities that truly create value. Only in this way can they be closer to PMF.

If you can't clearly answer why this task must be done by a robot or can't reduce the cost to a level that makes enterprise owners feel that not buying is a loss, then this so - called PMF is just an expensive self - sufficient game in the laboratory.

If you've done too much mental work, you can do some physical work. Spend three months in a factory workshop or observe in a nursing home for three consecutive days to see what the real pain points are and what jobs no one wants to do. This may be much more useful than adjusting parameters for a large model. After all, a good business plan is not achieved by gambling but by practice bit by bit.

After experiencing a whole year of the bustling embodied intelligence track, insightful investors have found that the Lovot in Japan invested by Zhu Xiaohu last year is actually the closest to the truth of PMF because it focuses on solving the need for a hug, and the market actually bought it. Looking back, the value of Zhu Xiaohu's investment is still rising. There must be a reason for him to invest only in Lovot and not in other projects.

In short, don't focus all your attention on the technological halo. First, figure out who (who) will really pay for it and why (why) they will pay - just like figuring out how many people are really willing to pay extra for autonomous driving.

This article is from the WeChat official account "Autoweek of Zhijian", author: Dada Wang. It is published by 36Kr with authorization.