What can a product manager do in the embodied intelligence industry?
From the poor experience of food delivery robots in apartments to the technical bottleneck in the stringing scenario of barbecue restaurants, two real cases reveal the current situation of missing closed-loop value for PMs in the embodied intelligence industry. This article deeply analyzes the dual challenges of technical ceiling and basic product design skills, discusses the conditions for establishing PM value and the PMF evaluation of current industrial scenarios, and points out the direction for practitioners.
I. Starting with two cases
During my entrepreneurial period in Suzhou, the apartment I rented was equipped with a food delivery robot from Yunji Technology. When I first moved in, I was quite looking forward to it - I didn't have to go downstairs to pick up takeout, which sounded very convenient. But the actual experience was completely different.
According to the assumption of product design, the process should be: the rider hands the takeout to the front desk, the front desk administrator puts the takeout into the robot and enters the room number → the robot takes the elevator autonomously → arrives at the door of the room and makes a call to notify the resident to pick up the meal → the resident takes it away. It sounds very reasonable.
However, in actual operation, this robot exposed a series of frustrating problems:
It has no hands to ring the doorbell. Although the apartment has a doorbell, the robot can only "wait" at the door of the room, and it cannot call the resident either - its own voice is very low, which is completely inaudible behind the closed door. Once I didn't hear the sound outside the door, it would wait for a while and return to the first floor with the meal along the original path. Several times I waited for half an hour, went downstairs and found that the robot had already returned.
It doesn't know "how many orders are there in front". After the front desk puts the takeout in order, I don't know the queuing status of my own takeout. The delivery experience is very poor. It may only take 15 minutes for the merchant to deliver the meal to the front desk, but the whole link from the first floor to the resident's room, including taking the elevator, corridor navigation, waiting, timeout return and re-delivery, may take 1 hour during peak hours. Later, I developed a habit: when ordering takeout, I have to confirm with the rider "how many orders are waiting in front", but many riders will put the meal at the front desk by default, click "delivered" and leave, and will not specifically call you to tell you "the meal has been put into the robot". When I wanted to check where my takeout was, I found that the takeout app already showed "delivered", but in fact the meal was still lying on the first floor.
My reaction at that time was: Did the product manager of this company ever go through the complete delivery process by himself?
The queuing visualization that can be solved with one App, the communication problem that can be solved with one push notification, and the waiting problem that can be solved with one reservation mechanism - these practices that have been verified countless times in the Internet industry are completely absent in the robot delivery scenario. If I were the PM of this company, I would definitely go through the user process first, and then make up for these basic experience points.
It should be noted that the robot at my residence may be an earlier batch of equipment, which does not represent the overall level of Yunji's current products. But the problems it exposed in my real experience for a whole year are essentially the arrears of basic product design skills - whether the PM really stands in the user's shoes and goes through the use case. This is not impossible to achieve technically, but no one has put their heart into it.
This case makes me feel that in the embodied intelligence industry, the PM role either does not exist, or does not play its role at all.
A couple of days ago, I went to the World Robot Conference. One of the barbecue restaurants I have shares in has stringing workers, side dish preparers and barbecue chefs. From the perspective of offline business, I took the stringing scenario to ask several robot manufacturers whether they can replace human workers with robots - their optimistic estimates are basically "5 to 10 years". I then found a senior embodied intelligence expert from a large factory and threw my question to him.
He did not answer directly, but asked me in return: "Are we trying to improve efficiency, or to attract foot traffic?"
I said I wanted to replace human labor. He said: "Robots can't do it, but automated meat stringing equipment can."
I pressed further: "Can the automated equipment take meat from the refrigerator, cut it, and string it by itself?"
"These are not possible," he replied very simply.
This case makes me feel that even if there is a PM and the requirements are defined clearly enough, if the technical level is not enough, the PM can do nothing.
Putting the two cases together makes me realize a problem: At the current stage, the embodied intelligence industry has not established the proper closed-loop value of PMs. On the one hand, no one is doing the basic work of product design, on the other hand, the technical ceiling limits the space for PMs to play their role. These two problems are different in nature, but they both point to the same conclusion - the PM role has not really gained a firm foothold in the industry.
II. Why the closed-loop value of PM has not been established
This feeling of mine does not come out of nowhere. Industry data and signals from the recruitment market are all confirming the same fact: At the current stage, PMs in the embodied intelligence industry have not really established their own closed-loop value.
First, look at the technical side.
Shandong Finance Network points out that embodied intelligence faces bottlenecks at the deployment level: the capability of offline models is difficult to directly enter the real closed-loop system, large models are difficult to run in real time on edge devices, and the whole industry is in the stage of moving from laboratory to industrial verification.
The data given by People's Daily Online during WRC 2026 is more intuitive: the current compliant data of real physical interaction scenarios in China is only 500,000 hours, while the commercialization and landing of robots requires tens of millions of hours of data, with a gap of more than 99%. Cao Peng, President of JD Cloud, said that the current robot industry is still on the eve of an outbreak and has not really broken out for a long time. The reason is that the generalization of the robot's "brain" is insufficient, and the root cause is the lack of real scenario data.
This precisely explains the expert's "can't do it" in the barbecue restaurant scenario - tasks like stringing meat involve long chains, non-standard operations that require tactile feedback, such as "taking frozen meat from the refrigerator → cutting different ingredients → stringing them", before the data bottleneck and generalization bottleneck are broken through, no matter how well the PM defines the requirements, they cannot be fulfilled. The manufacturers' answer of "5 to 10 years" may even be too optimistic about the technical level.
The existence of the technical ceiling directly limits the space for PMs to play their role. This is the first reason why the closed-loop value cannot be established.
Then look at the product side.
The problem exposed by the robot in the apartment is not that it cannot be achieved technically, but that the PM did not go through the user process. Queuing visualization, progress tracking, two-way confirmation - these practices that have been verified countless times in the Internet industry are completely absent in the robot delivery scenario. This is not a problem of C-end experience optimization, but a lack of basic product design skills. The current PM allocation in the industry is indeed concentrated on B-end scenario abstraction and technical boundary judgment, which is a reasonable choice determined by the technical level. But the problem is that even B-end-oriented PMs need to go deep into the front line to understand the real user process. Any PM can find the problem of the apartment robot as long as he lives in it and uses it for a month. But obviously no one has done this.
The lack of basic product design skills is the second reason why the closed-loop value cannot be established. PMs have not really stepped into the user scenarios.
III. Under what conditions can PM value be established
The previous analysis listed two reasons why PM value cannot be established. So what conditions are needed to establish it?
The core logic is: the value of PM is ultimately reflected in "the defined requirements can be technically realized, and the product can be accepted by the market". If the technology cannot realize the requirements (the barbecue restaurant case), or no one is willing to pay for the product (the apartment case), no matter how many requirements the PM defines, they are all in vain.
Therefore, to establish PM value, two prerequisites must be met at the same time: the technical level can cover the requirements, and the product can run through the market.
China Information Weekly gave a measurement standard in its WRC 2026 report - to cross from Demo to productivity, four barriers need to be passed: First, real demand, the product solves the rigid demand that customers are willing to pay for; Second, usable product, with stability, efficiency and safety reaching the standards of the production environment; Third, economically feasible, customers can calculate clear ROI, and suppliers can also get reasonable profits; Fourth, scalable and replicable, deployment does not rely on a large number of customizations.
The essence of these four barriers is to unify "what technology can do" and "what the market can accept". Only when all four barriers are passed can the "requirements" defined by PM be truly implemented, and the value of PM be established. If any barrier is not passed, the PM's efforts will get stuck halfway - either the technology cannot produce the product, or no one buys it after it is produced.
But there is an important supplement here: Even if not all four barriers are passed, the value of PM in the basic skills of product design can still be established first. The apartment case shows that within the reach of existing technology, improving interaction design and optimizing user experience can significantly improve the usability and satisfaction of the product. This part of the value does not need to wait for technological breakthroughs, and PMs can realize it right now.
Tang Jian, CTO of Beijing Humanoid Robot Innovation Center, put forward a three-stage path for the industrial landing of humanoid robots at WRC 2026: industrial and special scenarios achieve breakthroughs first, commercial service scenarios accelerate penetration, and family and generalized scenarios achieve ultimate landing. This echoes the framework of the four barriers - the earlier the scenario, the closer the technical level is to the range where PMs can give play to product value; the later the scenario, the more PMs need to wait for technological breakthroughs.
IV. Current industrial status: PMF evaluation of three scenarios
Based on the above framework, I will evaluate the main landing scenarios of current embodied intelligence one by one, to see in which scenarios the PM value has been established and which have not.
Scientific research, education and industrial logistics: the closest to running through the four barriers.
During WRC 2026, Xi Yue, co-founder of Starbot Era, said that as the industry's first embodied logistics solution that announced the completion of PMF verification, it has achieved normal operation with SF Express and China Post in more than 10 logistics centers across 5 provinces and cities, and the operation efficiency in some scenarios has exceeded that of humans, and the scenario migration speed has been increased from two months to within one week.
The PM value in this scenario is the closest to being established, but the threshold is extremely high. PMs need to understand both technical boundaries and industry cognition - scientific research and education customers buy "quantity", while industrial logistics customers calculate "efficiency", and the PM playbooks for the two are completely different. The latter is exactly the part that cross-border PMs find the hardest to make up for at present.
Commercial services: clear demand, but neither product design nor economic model has been run through.
The Yunji robot in the apartment is a typical example. The demand is clear, the technology can run the Demo, but the lack of product design leads to reduced experience. In this scenario, the "real demand" is established, but "usable product" and "economically feasible" have not been run through. Part of the problems on the product usability level can be solved through better product design - this is exactly the direction PMs can start working on at the current stage.
Home services: immature technology, vague demand.
Tang Jian listed "the ultimate landing of family and generalized scenarios" as the final stage of the three-stage path. No investment should be made in this direction at present.
V. What should embodied intelligence PMs do
Within the reach of existing technology, make the product experience solid - that's what the Yunji case taught me.
In scenarios where the technology has not yet reached, break down the requirements into verifiable single points - that's the thinking brought by the barbecue restaurant case.
The former tests the basic skills of product design, while the latter tests the judgment of technical boundaries. Neither of them needs to wait for technological breakthroughs, and can be started right now.
Back to the robot in the apartment: queuing visualization, progress tracking, two-way confirmation - the lack of these basic experiences is not because the technology cannot do it, but because no one has gone through the user process. If the PM is willing to spend a week deep into the front line, design a simple App to do these things, the user experience will be qualitatively improved. This is a dimension that almost no one in the current industry touches, but it is exactly the part that PMs should make up for the most.
Back to the barbecue restaurant scenario: for meat stringing, manufacturers say "5 to 10 years", and the expert says "robots can't do it". But what if we break down the long-chain task of "stringing meat"? Taking meat out of the refrigerator, putting it on the cutting table, cutting the meat, stringing it - each step has different technical difficulty. Maybe the single point of "taking meat out of the refrigerator and putting it on the cutting table" can be achieved with existing technology. This is what PMs need to do when the technical level is not enough: instead of defining a large and comprehensive solution, break down the scenario into single points that can be technically verified, and then clearly define the acceptance criteria for success rate, efficiency and cost. The large factory expert's rhetorical question "Are we trying to improve efficiency, or to attract foot traffic" is actually a reminder: if the goal is to replace human labor, then we need to find the part that technology can fulfill, instead of expecting to achieve it in one step.
So my answer is: what embodied intelligence PMs can do is not to make promises where the technology cannot reach, but to make the product closed-loop solid within the reach of existing technology, and find the safety margin of the technical level through scenario decomposition.
The technical level is a hard constraint, and product design is a soft arrears. What PMs can change is the part of soft arrears; what PMs must respect is the part of hard constraint.
References
[1] People's Daily Online: Industry Observation: How can embodied intelligence quench its thirst for data?
https://finance.people.com.cn/BIG5/n1/2026/0822/c1004-40784264.html
[2] China Information Weekly: 2026 World Robot Conference opens: Embodied intelligence ushers in a key window period for large-scale application
http://mp.weixin.qq.com/s?__biz=MzkxNzU5ODUzMg==&mid=2247569468&idx=1&sn=b924474e63ed6090a00f2e9a24dee03e
[3] NetEase: Live WRC 2026: Xi Yue from Starbot Era: PMF has been verified in logistics scenarios, and the large-scale landing of humanoid robots will focus on toB first
https://www.163.com/dy/article/L51RMOOI05199NHJ.html
This article is from the WeChat official account "Everyone is a Product Manager" (ID: woshipm), author: nathan, published with authorization from 36Kr.