Original reporting by Unique Research · September 7, 2026, 16:00 · Shanghai
AI Industry Watch
What the robotics industry lacks most today may not be intelligence. It may be a business case that actually adds up.
Let us start with a fact.
Today, you can spend 100,000 yuan on a high-degree-of-freedom dexterous hand, and it might break in ten days. With better luck, it might last a month or three months, which already counts as a long life. And repairing it is not necessarily a matter of tightening a screw. An entire finger may have to be taken apart and reassembled, with the repair ultimately costing more than simply replacing the hand.
That was what Zhang Long of DexRobot said during a panel discussion. He offered this analogy:
“Today’s dexterous hands are like the hand of someone with Alzheimer’s disease.”
The hand itself, in other words, is not the main problem; the brain has not yet become fully functional. Whether through VLA or a world model, the entire industry is waiting for a breakthrough in models. At least for now, that moment has not arrived.
The panel focused on “diverging application scenarios.” Its four participants were Hu Xiaoping of Flexiv, Zhang Long of DexRobot, Li Jufan, formerly of Lymow, and Dou Haotong of Hechuang Shengwu. Their fields spanned industrial arms, dexterous hands, yard robots and biological laboratories. And all four were working through the economics.
How do you calculate the economics of essential demand, a demonstration installation or volume delivery? How do you compress a nonstandard solution into a standardized product? Put plainly, what the robotics industry lacks most today may not be intelligence. It may be a business case that actually adds up.
When commercialization in robotics comes up, many people’s first instinct is to look for high-frequency, low-cost applications: machine loading and unloading, sorting and material handling, where volume spreads the cost.
Zhang’s assessment is that applications capable of supporting commercialization at this stage may not involve putting robots into direct competition with the lowest-cost labor. Instead, companies should first look for high-value workers.
Take high-containment biological laboratories. People qualified to enter them are at least PhDs or postdoctoral researchers, with annual salaries of one million or two million.
Each time they enter a sterile environment, putting on protective clothing takes half an hour to an hour. After working for a while, if they need the bathroom, they must come back out, take everything off and put it all on again. Even with a decontamination shower, there is still a risk of contaminating the experimental environment—not to mention hazardous environments involving radiation or infection risks.
Lay out the calculation: on one side, a postdoctoral researcher earning one million to two million a year, with limited effective working hours each day; on the other, a robotic arm costing around 200,000 that can work around the clock, seven days a week.
Even if the arm does nothing more each day than go in, pick up a test tube, dispense solvent and return it to the refrigerator, the economics can still make sense.
“The value of an application does not depend on how complicated the movement is. It depends on the safety costs and the cost of high-value labor behind it.”
Hu Xiaoping of Flexiv added another angle. Robotics contains a rather harsh paradox: many things people consider effortless, such as packing a cardboard box, are extremely difficult for robots. Conversely, moving along a precise one-meter trajectory is difficult for people but easy for an industrial robot. Their capabilities are structured very differently.
That is why Flexiv has chosen human-like force control: allowing a robot to adapt on its own to changes in the environment and variations in error.
Where has this been deployed? Not in glamorous consumer-electronics factories, but in food processing. Cutting meat and handling chicken or fish involve different shapes, materials and temperatures—problems automation previously struggled to crack.
Now the sector is struggling to recruit workers. Its workforce is aging, while younger people would rather deliver takeout than perform highly repetitive work in cold, damp workshops.
Hu gave an example: if you take a China Eastern flight from Shanghai, some components of your in-flight meal may have been assembled with the involvement of Flexiv robots.
Li Jufan, formerly of Lymow, works on consumer yard robots. His opening was candid: compared with the previous two speakers’ products, his category involves less advanced technology and uses less AI.
Yet precisely these “AI 1.0-era” products commercialized first. The global retail market for robotic lawn mowers has already exceeded US$1 billion, and penetration of robotic pool cleaners in North America passed 20% two years ago.
Why were these the first two categories to succeed? Li posed a question: after buying a robot vacuum, do you still have a broom at home? A mop?
Of course you do. A robot vacuum does not replace the tools. It replaces the user’s time.
A lawn comes with a whole list of jobs: mowing, edging, clearing fallen leaves, watering, fertilizing, pest control and pulling weeds. Which should be robotized first?
Li’s screening framework has just four criteria: is the task frequent enough, is the pain point sufficiently painful, can existing technology solve it, and will users accept the cost?
Mowing happens once a week, but the pain depends on the person. Some find it therapeutic. Others have gone ten years without buying any mowing equipment and will pay someone else to avoid it.
Pool cleaning is tougher: first stir things up, then scrape the pool walls, then work through the pool bit by bit with a manual vacuum. Almost nobody enjoys it, and it is also a weekly task.
By contrast, everyone in user research says pulling weeds is painful. But today’s combination of vision and mechanical manipulation cannot reliably spot a weed and precisely pull it out. However painful the problem, if the technology cannot reach it, it must wait.
Three fields, one conclusion: essential demand is not found by following the excitement. It is the intersection revealed by doing the calculation.
After finding essential demand comes a harder hurdle: turning a demonstration installation into actual orders.
Hu first has to contend with an intensely competitive market. A domestic collaborative robot might sell for 20,000–30,000 today, 20,000 tomorrow, and perhaps less than 10,000 the day after. Flexiv’s adaptive robots require substantial R&D investment and cost more. A customer’s first question is: why are you so expensive?
He has two approaches.
First, win leading customers in automotive and consumer electronics—NIO, XPeng, Li Auto and Tesla. Once a reference customer is established, others will follow.
Second, do not calculate only “how long it takes one arm to pay for itself.” Calculate the complete solution. A flexible arm may replace an entire part of a process, reduce the need for custom equipment, improve production-line flexibility and lower maintenance costs.
The most convincing evidence comes from food processing. Margins there are already thin. If even a low-margin industry is willing to buy in volume, the calculation genuinely works.
For Zhang, the calculation moves to another level. He sees a fundamental change taking place in industrial demand.
Mass production used to answer the question of availability: did you have a car to drive or a phone to use? Now that people have cars, the question is what kind of car to drive and which configuration to buy.
Orders are becoming more fragmented. Many component orders contain only 200 units, or 300–500. You cannot build a new production line for 200 products.
The economics of dexterous manipulation therefore cannot be reduced to “how many workers it replaces.” Zhang says the right question is: how much additional business can I take on because I have this system? A line that can do the work previously done by five lines and handle ten different kinds of orders supports a different calculation.
But Zhang acknowledges that it is too early to apply that calculation across the industry. Embodied AI currently faces three shortages.
It lacks data, the scarcest and most expensive resource in embodied intelligence. It lacks collaboration across robot embodiments: your hand and mine each collect their own data, and neither can use the other’s. And it lacks interface standards: every company’s hardware interfaces are different.
In his words, the industry is nowhere near its “USB-C moment.”
During the panel, Li raised a question on the spot: large language models produced OpenAI and Anthropic. Could robotics likewise produce a foundation model spanning every kind of robot embodiment?
Zhang said he had worked at Baidu and knew how large models were built by pouring money into them—from tens of millions to hundreds of millions, then billions. An embodied foundation model would require even more investment.
Given the size of today’s embodied-AI startups, if they have to build hardware, run their core businesses and also spend tens of billions on foundation models, everyone may burn through their resources together—and fail together.
He expects the companies that eventually build world models to be today’s large-model developers. Embodied-AI companies would contribute data, application environments and robotics know-how, then fine-tune on the larger companies’ foundations.
When the conversation turned to consumer products, Li brought it straight back down to earth. Many hardware startups begin with crowdfunding. But he was clear: crowdfunding is advance marketing and product-market-fit validation. What it is not is commercialization.
Apple or Huawei can hold a launch event and put a product on sale a week later, because the brand itself brings traffic. Startups do not have that privilege. They have to explain their product concept three to six months ahead, slowly building interest from zero. However impressive the crowdfunding total, it shows only that people are willing to buy into the concept.
The real trial comes in mass production. Structural components, electronics and the supply chain are all new. He has seen early defect rates above 50%. What is the greatest fear? As he put it, the product envisioned in the presentation is a young woman, but the mass-produced version turns into a middle-aged auntie. Then you are finished.
The first job after mass production is not glamorous either: make sure after-sales support can handle the load. Build technical support, customer service, the knowledge base and FAQs in advance, working backward from the delivery date.
In his most recent venture, he oversaw the entire customer-facing operation. When products first arrived and problems emerged in the community, what could they do if there were not enough staff? Everyone became customer service. They made a roster, with someone available from 8 a.m. until midnight.
A prominent negative review in a Facebook Group, with 30–50 likes, can knock down your conversion rate the next day. The early pool of users is simply that small.
His original wording was rough, but to the point:
“Those few thousand early adopters from crowdfunding are your angels. Treat your angels well and you will be rewarded. Treat them badly and there really will be consequences.”
Dou Haotong of Hechuang Shengwu brought the discussion into life sciences. His company received US$200 million in investment from SoftBank during the pandemic and is now working with NVIDIA on embodied intelligence.
In his view, the machines in a biological laboratory are not fundamentally different from robotic lawn mowers. They are not even necessarily “fully AI.” Much of the work consists of repeatedly transferring liquid from large tubes to smaller ones.
It is rather like cooking: chopping, grinding and heating, adding some “seasoning”—biological reagents—and finally putting it into a PCR machine.
He offered a particularly vivid explanation of throughput.
If ten friends come over at the weekend, you do not need two chefs; you need a bigger pot. If a village banquet has 1,000 guests, you do not need 1,000 chefs either. You need one head chef, five to ten assistants chopping vegetables, a few big pots and shovels to stir-fry the food.
Specialized machines handle throughput. Robotic arms and hands do work that people cannot do, or do it while people are asleep. The AI brain handles thinking. These are three different things.
Another image made a strong impression on the audience: in many laboratories today, a robotic arm moves in front of an instrument to “press a button.” It is not that the arm lacks intelligence. The machine is too old: it has buttons but no AI interface. An arm has to be sent to press them in place of a person, rather like starting a tractor with a hand crank.
Asked which matters most in winning an order—throughput, accuracy or traceability—Dou gave a somewhat unexpected answer: traceability. The life-sciences industry’s insistence on research integrity means that every step of an experiment must be reconstructable. Efficiency can be compromised; traceability cannot.
The final question went to everyone. Robotics products all contain some degree of customization, but volume delivery cannot mean rebuilding the product for each customer. How do you reduce that nonstandard element as you move from a demonstration installation to large-scale orders?
The four answers are especially interesting together because they are entirely different.
Hu’s approach is to move upstream: find a technical problem shared across industries and solve it with the same core technology. Standardize the core, then leave the rest to peripheral equipment and local adaptation.
Zhang’s approach is to cut back: limited intelligence within a limited space.
A hand may theoretically do 1,000 things, but at a particular workstation it may only need to perform ten of them frequently and well. The other 990 functions might be used twice a year, yet require ten or twenty times the cost.
In industrial settings, where ROI is scrutinized to the limit, that does not make sense. Cut what needs to be cut.
Li put it more directly: find the greatest common denominator. Once a consumer product reaches scale, it follows the familiar consumer-electronics playbook. Just as phones are divided into price tiers, standardize core capabilities such as cutting and slope handling, then create different SKUs for lawn size, gradient and purchasing power.
Dou’s approach is the most “data-driven.” Around 15,000–16,000 biological laboratories worldwide use his company’s equipment, including some of the world’s leading life-sciences laboratories.
With large amounts of real usage data accumulating, it becomes clear what people are actually doing with the equipment. The common denominator can be inferred from that data.
Find common ground, subtract functions, segment SKUs, mine data. Four routes lead to the same destination: standardization is not simply designed into existence. It grows out of real applications.
As the panel drew to a close, Dou responded to Zhang’s remark that the “GPT moment has not arrived.”
He said he had spent the previous six months consistently using Pony.ai’s robotaxis to commute and meet friends. Sometimes the car is slow and not especially convenient, but he keeps using it.
It has given him a strong impression: this kind of technology does not need to astonish everyone on one particular day. It only needs to add vehicles one by one and expand coverage area by area, gradually becoming part of everyday life.
Ordinary people may not even realize that it is already entering society.
In his words:
“Physical AI’s GPT moment may not arrive suddenly on a particular day. It may arrive gradually.”
Before that day comes, the companies that do best in this industry will probably be those with the clearest grasp of the economics—even if their models are not the most dazzling.
Guests
Hu Xiaoping, Vice President, Flexiv
Zhang Long, Head of Government Relations and Ecosystem, DexRobot
Li Jufan, former Partner and CMO, Lymow
Dou Haotong, Head of Asia-Pacific, Hechuang Shengwu
Moderator
Kang Zhengzhong, Partner at GreenSeed Global / Head of Unique Research Shenzhen
Kang Zhengzhong: Our first panel this afternoon is about “diverging application scenarios.” Robotics is very hot right now, but how do you find a task frequent enough to make a robot an essential solution? That is our first question.
Kang Zhengzhong: Second, once you have found that essential demand, what calculation do you need to make to turn a demo into a real reference installation? Third, after the first order from that installation, how do you keep replicating it and win orders at scale? We will discuss those three perspectives. First, could the four of you briefly introduce yourselves and your companies’ businesses?
Hu Xiaoping: I am Hu Xiaoping from Flexiv. I gave a presentation earlier. Our core focus is improving robots’ manipulation capabilities. By enhancing what robots themselves can do, we hope to empower different industries.
Zhang Long: I am Zhang Long from DexRobot. We focus on dexterous manipulation, including data collection at the front end, our own models and five-fingered dexterous hands with many degrees of freedom.
Zhang Long: As you look at robots, you increasingly realize that doing the actual work—solving the last step—may ultimately still depend on the “hands.” We therefore focus more on solving practical manipulation problems.
Li Jufan: I am Li Jufan. My career has included experience within Huawei, and I have also incubated new product categories within Ecovacs and Dreame. Most recently, I became an entrepreneurial partner at Lymow. My work has largely revolved around consumer electronics and consumer robotics.
Dou Haotong: I am Dou Haotong from Hechuang Shengwu. We mainly work on robots for biological experiments. During the pandemic, the company completed a US$200 million financing round backed by SoftBank. We are now also collaborating with NVIDIA on some embodied-intelligence work.
Kang Zhengzhong: Let us start with Mr. Hu. Traditional automation has typically required stable environments and repetitive workstations. Yet the places that genuinely lack workers are often precisely those with extensive variation, where earlier automation could not solve the problem.
Kang Zhengzhong: From your perspective, which tasks are both frequent enough and sufficiently complex to require an adaptive robot? And when industrial customers adopt adaptive robots, what is the first calculation they make—labor costs, quality risks or something else?
Hu Xiaoping: Flexiv initially took a more technology-driven or product-driven approach, rather than being entirely demand-driven. Our first question was whether a technical breakthrough in robots’ own capabilities could push their generalization and versatility further.
Hu Xiaoping: From the beginning, we therefore chose human-like force-control capabilities, allowing a robot to adapt autonomously to environmental changes and variations in error during manipulation. That is why we call it an “adaptive robot.”
Once a robot’s own capabilities improve, you discover that many more applications become possible at once.
There is an interesting paradox in robotics. Many things that people do easily are difficult for robots. Take packing a cardboard box: a person wonders what could possibly be difficult about it, but a robot finds it extremely difficult. Conversely, some things that people find difficult, such as following a very precise one-meter trajectory, are easy for an industrial robot.
The structure of their capabilities is therefore very different.
But people can solve a tremendous range of problems. If we can give robots those human capabilities, the potential market is very large. This is what we have been working on all these years.
When it comes to specific applications, food and agriculture have actually made the deepest impression on me.
Automation penetration in this field is still relatively low. It remains a classic labor-intensive industry. And its problems are becoming increasingly apparent: workers are getting older, recruiting and managing people cost more, and younger people are increasingly unwilling to perform highly repetitive work, sometimes for extended periods in cold, damp environments.
They may prefer an occupation offering more freedom to working in such a factory.
This is therefore a field urgently in need of an automation upgrade, with very high task frequency. Food is also closely connected to everyone’s daily life. A food-safety problem can even be fatal to the business.
Why has food automation been difficult? Food has a characteristic variability: different shapes, different materials, different environments and different temperatures.
Cutting meat, handling chicken or fish, or processing different ingredients requires a manipulator with the adaptability of a human hand.
We have already deployed in many such applications, including meat, chicken and fish processing. These systems are now running normally on production lines.
Another example is mainly in Shanghai, where our headquarters are located. If you take a China Eastern flight from Shanghai, some of the meals served on board may have been prepared with Flexiv robots participating in their assembly.
I think this is meaningful work, and it has already produced some good results.
As for what customers ultimately consider, the first step is always the technology: can you actually do it?
If you cannot do the job, ROI is meaningless.
Once it is technically achievable, calculate the ROI. That determines whether you can complete the last mile to deployment.
Kang Zhengzhong: Mr. Zhang, people’s most immediate impression of DexRobot may still be its dexterous hands. But as I understand it, you are advancing across hardware, data, models and applications.
Kang Zhengzhong: Humanoid robots are very popular right now, but whether someone orders one does not depend on whether it has wheels or two legs. What matters most is whether it can work. So, first, what high-frequency tasks must dexterous hands solve to be more than a “beautiful accessory” on a robot? Second, if the customer is a research institution or robot manufacturer, what do they actually examine when accepting a hand—degrees of freedom, service life or maintainability?
Zhang Long: The change is very apparent.
In 2025 in particular, many hand manufacturers would say, “We make hands,” placing special emphasis on a high number of degrees of freedom. By 2026, practically no company describes itself as only making hands. Everyone has started extending into “manipulation,” including data, models and more.
Why?
Think about it: what is one of the parts of the human body most capable of creating value? The hand.
We therefore do not simply define the hand as a core component. There is an important underlying idea here: do not reduce the hand to a tool.
The grippers, paddles and forks we used in the past are essentially tools. Tools have specific situations in which they are used.
A human hand is different.
The versatile, generalizing hand we envision should be able to use many kinds of tools. It is not itself the tool.
“Human-like” therefore means more than looking like a human hand. More importantly, people use their hands to operate tools; they do not accomplish everything with their bare hands alone.
Now let us talk about high-frequency applications.
Frankly, many applications we see that can genuinely support commercialization do not necessarily make their case through “high frequency and low cost.” It may be better to start with high-value applications.
Hazardous environments and scientific research, for example.
Someone here also works on robots for biological laboratories. A person qualified to enter a high-containment biological laboratory is often at least a PhD or postdoctoral researcher. That person’s labor is already expensive.
Entering a sterile environment may require half an hour or even an hour to put on protective clothing. After working for a while, if they need the bathroom, they must come out, take it off and put it on again. Even after a decontamination shower, they may still contaminate the experimental environment. In hazardous environments involving radiation, for example, there is also the risk of infection or injury to people.
The value of such an application therefore does not depend on how complicated the movement is. It lies in the safety costs and high-value labor costs behind it.
Compare a postdoctoral researcher earning one million or two million a year with a robotic arm that might cost 200,000 and work around the clock, seven days a week. Even if it only collects test tubes, dispenses solvent and puts them back in the refrigerator each day, the economics may still make sense.
That is entirely different from replacing an ordinary worker earning a few dozen yuan per hour.
At this stage, then, I honestly do not think dexterous hands are as good in real large-scale manufacturing as people imagine.
Rather than fixating on loading and unloading, sorting or material handling and competing directly with the lowest labor costs, we might first target high-value fields such as safety and scientific research, replacing some high-value work while protecting people.
Today’s problem depends not only on “how well the hand is made,” but also on the world models beyond the hand itself.
What are today’s dexterous hands a little like? The hand of someone with Alzheimer’s disease.
The hand itself may be fine. Once the person’s brain is functioning normally, the hand immediately works properly. But, unfortunately, the brain is not yet that clear-headed.
Whether through VLA or a world model, everyone is waiting for that model breakthrough. At least for now, it has not arrived.
When a customer accepts a hand today, I therefore think maintainability is especially important.
Almost every high-degree-of-freedom dexterous hand today is essentially a consumable.
Spending 100,000 yuan on a hand does not mean it will work reliably for a year. Ten days, one month or three months may already count as a long time. With many dexterous hands today, if one finger breaks, you cannot simply remove a small screw and fix it. You may have to remove the entire finger, or even replace the whole assembly beneath it. The repair may ultimately cost more than replacing the hand.
This is a problem high-degree-of-freedom dexterous hands have to face once they enter the real market.
Why are there still so many compromise solutions in industrial applications? People continue to use grippers or three-fingered designs.
Performance, ROI and stability form a triangle in which all three cannot be maximized at once. For now, we can only strike a balance.
Of course, as costs fall, stability improves and world models advance, we still look forward to more general-purpose hands genuinely transforming manufacturing lines.
Today’s factories still follow an underlying logic established decades ago.
Agents have already begun transforming office work. I think factories should eventually move to a new foundation driven by models and agents too. The hardware you need will then change. It may be intelligent hardware, represented by dexterous hands, that works within limited spaces and with limited intelligence.
Kang Zhengzhong: So we can understand this to mean that embodied intelligence’s real “GPT moment” has not yet happened.
Zhang Long: Correct. Our current assessment is that it has not.
Kang Zhengzhong: You cannot begin by expecting to get everything right in one leap, building a general-purpose robot that solves every problem.
Kang Zhengzhong: Mr. Li has worked on yard robots, a typical category of service robot. Yards are actually very complex environments. European yards may be smaller with clearer boundaries, while North American yards are larger and have fewer physical boundaries. Grass varieties, slopes and weather also keep changing.
Kang Zhengzhong: Based on your years of experience, what tasks are best handed over to ground-based robots first? Why have robotic lawn mowers and pool cleaners been the first to succeed?
Li Jufan: Our category differs somewhat from what the previous two speakers do. Frankly, it involves less advanced technology and is more commercialized.
If we talk about AI today, robot vacuums, robotic pool cleaners and robotic lawn mowers all use relatively little of it.
“AI” might mean that I do not mow along a completely hard-coded route, or that I use some AI in camera-based visual recognition to identify a hedgehog, a slope, a hole or a barbecue grill.
Although we also call them robots, they are more like products from the AI 1.0 era.
Precisely because of that, these categories commercialized earlier. The global retail market for robotic lawn mowers has now exceeded US$1 billion.
So which applications are best suited to commercializing first?
Let me first ask everyone a question: do robot vacuums and robotic lawn mowers replace manual tools?
After you start using a robot vacuum, do you still have a broom at home? A mop?
Of course you do.
What the robot actually replaces is not the tool but the user’s time.
That is important.
Returning to “high frequency,” however you describe a consumer business, the starting point is still the user and the application.
A lawn involves many jobs: mowing, edging, clearing fallen leaves in autumn, watering, fertilizing, loosening the soil, pest control and weeding.
Which of these is worth robotizing first?
First, look at frequency.
Mowing is usually weekly. In autumn, collecting leaves may also be weekly. Some robotic mowers now add a basket to collect fallen leaves at the same time.
Tasks such as pulling weeds or trimming hedges are much less frequent.
Second, consider how much mental and physical effort the task takes, and whether the user actually enjoys doing it.
Mowing is a little tricky in this respect.
We have met users who really enjoy it. They like the smell of the grass, find it relaxing and feel a therapeutic satisfaction afterward.
But another group hates it.
Some say, “I have not bought any mowing equipment myself in ten years. I hire someone because I think mowing wastes too much time.”
Another kind of user may not dislike mowing, but has no time for it.
In the United States, for example, an acre of lawn may take two hours a week, or eight hours a month. That is already a substantial time cost.
When you cross-match those conditions, mowing turns out to be a task well suited to replacement by a robot.
Pool cleaning is an even clearer example.
To clean a pool, you first stir up what is inside, scrape the dirt on the walls down to the bottom, then use a manual vacuum to work through the pool floor bit by bit.
Almost nobody enjoys it.
Pool cleaning is also a typical weekly task.
Two years ago, robotic pool cleaners already had penetration above 20% in North America. Why? Because the task is frequent enough and painful enough.
At the most fundamental level, then, it comes back to user needs and applications.
Next comes a second question: can the upstream technology produce the product?
A robotic mower is relatively simple: fit a blade and a miniature autonomous-driving system, and it can operate.
With suction or adhesion, positioning and navigation, and route planning, a pool robot can also generally operate—though particularly complicated pools are harder.
Some tasks users really dislike are still beyond the technology.
Weeding, for example.
If a few weeds appear in the lawn, I first have to identify them accurately, then precisely spray a little chemical treatment or pull them out directly.
Our user research found that many people particularly dislike weeding.
But today’s vision and mechanical manipulation have not yet reached the point where a robot can see a weed and pull it out as precisely as a dexterous hand would.
If that capability genuinely matures one day, I think users will certainly buy it.
So which applications are suitable for robots? The central questions are these:
Is the task frequent enough? Is the pain point painful enough? Can existing technology solve it? And can the product ultimately be made at a cost the user accepts?
It is essentially the intersection of those conditions.
Kang Zhengzhong: Finally, a question for Mr. Dou. Many processes in life-sciences laboratories look very repetitive. But automating them turns out to be more complex because reagents, experimental protocols, data and traceability impose strong constraints.
Kang Zhengzhong: Which high-frequency processes are suited to robots? When customers adopt laboratory robots, what matters more: reducing human error, increasing throughput or making results traceable?
Dou Haotong: Let me continue with the robotic lawn mower example.
People imagine that experiments conducted by PhDs or members of scientific academies must be very sophisticated. In reality, our machines are somewhat like robotic mowers. They are not “fully AI” systems either. Even today, the part that genuinely uses AI may mainly be vision.
I previously worked in 3D printing. More than ten years ago, I was among the relatively early 3D-printing entrepreneurs in China.
Why did I later enter this industry? Partly because I had accumulated a great deal of motor-control technology.
Consider CNC machines. A few years ago, mentioning CNC generally meant going to a factory in Dongguan to see a huge machine costing hundreds of thousands. Now you can buy a small CNC machine for home use for 20,000 or 50,000 yuan, or a 3D printer for a few thousand yuan.
You may now think of it as an “intelligent machine,” but just a few years ago it was simply a machine inside a factory.
Laboratory equipment is similar.
Many of our devices look much like 3D printers, but their work resembles cooking at home.
What happens in biological experiments? Cells are chopped up, ground, heated and “cooked.” Then some seasoning is added—we call it a biological reagent—to produce a chemical reaction, before the material finally goes into a PCR machine.
During the pandemic, the COVID tests people saw were essentially tests for the virus, ultimately determining whether a sample was positive or negative.
I therefore often tell people within our company to distinguish several things clearly:
There is the brain in the cloud—AI. There is a humanoid machine, which may look very human and have twenty or thirty joints, or even more than forty, but is still fundamentally a machine. There are robotic arms and hands. And beyond those are specialized machines for vacuuming, weeding, cooking, heating and other tasks.
A machine is a machine. What makes it intelligent is AI.
But AI also has things it cannot do today. That is precisely the problem facing physical AI.
For example, some of our current work with NVIDIA is essentially digital-twin work: scanning things from the physical world and simulating them in a virtual world. We first simulate online whether an experiment can succeed, then have the real machine perform it once the simulation succeeds.
But if we only ever simulate in a silicon-based world, scientific development may still be limited.
In my personal view, the real breakthrough lies in whether we can connect robotic arms, robotic hands, robot vacuums, robotic mowers and laboratory machines together.
One product type I particularly like, for example, adds a small robotic arm to a robot vacuum. It does not just clean the floor; it also picks up LEGO pieces and other small objects.
Returning to the moderator’s question about throughput:
A great deal of work in biological experiments involves someone repeatedly transferring liquid by hand from a large tube into smaller ones.
It is like mowing.
Suppose you have two acres of lawn. If you do not use a mower and instead have a humanoid robot cut it a small patch at a time, how many humanoid robots must you buy? A hundred?
Biological experiments are similar.
Some companies in this industry had very high valuations when they went public, only to fall quickly afterward. Was it capital-market hype, or was the real demand insufficiently grounded? That question is worth considering.
At home, one person can buy ingredients and cook. If ten friends come over at the weekend, do you necessarily need two people cooking? Not necessarily. You are more likely to need a bigger pot.
If you get married and host a village banquet with 1,000 guests, do you need to hire 1,000 chefs or buy 1,000 humanoid robots?
No.
You may need only one head chef, five to ten assistants chopping vegetables, and a few enormous pots, using shovels to stir-fry the food.
Specialized machines, humanoid machines and AI brains therefore address business processes in different ways.
The AI brain does the thinking.
Humanoid robots, arms and hands can replace many operations that are difficult for people, or keep working while people sleep.
But specialized machines are what actually increase throughput in many tasks.
The problem is that a specialized machine with no AI at all also struggles to keep increasing its throughput.
Why are many laboratories now accepting robotic arms and hands? You will see an arm moving up to an instrument to “press a button.”
It is not because the arms or hands are insufficiently intelligent. It is because our machines are too old.
They have only buttons, with no AI interface at all.
So an arm has to go over and press the button in place of a person, rather like a hand-cranked tractor.
If those machines are all upgraded in the future with interfaces through which AI can control them, cloud AI could directly control weeding robots, robot vacuums, robotic hands and arms, and laboratory equipment. Only then might we truly connect the entire workflow and substantially increase throughput.
Kang Zhengzhong: In the first round, we identified relatively frequent tasks with essential demand in different fields. In the second, let us discuss a more practical calculation: how do you turn a technical solution into a real demonstration installation? Let us start with Mr. Hu.
Kang Zhengzhong: Industrial customers clearly will not pay simply because “this robot is intelligent.” When buying a robot, they ask about yield, cycle time and the payback period. Adaptive robots are also relatively new and generally more expensive than traditional solutions. How do you demonstrate their value—less labor, fewer line changeovers, lower maintenance costs or something else?
Hu Xiaoping: That is a crucial question. It really determines whether the business can be deployed.
An adaptive robot is a relatively new product with higher R&D investment, so pricing is inevitably an issue at the start.
This is especially true given the intense competition in China’s industrial and collaborative robot markets.
A collaborative robot might cost 20,000–30,000 today, 20,000 tomorrow and possibly less than 10,000 the day after.
A higher-priced new product will naturally face challenges in that market.
Customers will inevitably ask: why are you so expensive?
Many customers also did not previously know that force control could solve certain problems. They still think in terms of traditional industrial-robot solutions and naturally bring that older pricing framework into the comparison.
We therefore knew from the outset that this challenge would exist.
Our first approach is to find relatively leading companies.
In automotive and consumer electronics, we first work with industry leaders—for example, NIO, XPeng and Li Auto in China, and international companies such as Tesla.
If those manufacturers accept your product, they naturally become industry reference cases, making other customers more willing to try it.
Second, we emphasize the value of the overall solution.
Deploying a flexible robotic arm does not necessarily solve just one process step. It may address an entire section of production.
Installing that robot may mean less custom equipment, a more flexible line, lower ongoing maintenance and operating costs, and fewer workers.
We therefore do not simply calculate “how long one arm takes to pay for itself” with the customer.
We calculate the value created by optimizing the entire solution.
After years of exploration, we have also achieved volume deployment at some food-processing companies.
As everyone knows, many food processors have thin margins. If a low-margin industry is ultimately willing to adopt our solutions in volume, that already shows that the value and cost-effectiveness can support the business case.
Kang Zhengzhong: Mr. Zhang mentioned two kinds of customers: research institutions and customers working on real operational tasks. Their budgets and acceptance criteria differ considerably. Which group currently represents higher-quality customers?
Kang Zhengzhong: Also, people previously might have “bought a dexterous hand for a demonstration.” Now some are buying for particular tasks. What common standards does the industry still lack in this transition?
Zhang Long: At this stage, research organizations and manufacturers with fragmented, flexible requirements tend to have greater demand for dexterous manipulation. Of course, it does not necessarily have to use five fingers.
Research has a particular characteristic: it is not solving today’s problems, but tomorrow’s.
I want to know where the true technological frontier lies.
Whether we call it embodied intelligence or physical AI, I believe methods of production will certainly change over the next five or ten years.
When research institutions buy the most advanced hardware and software today, they are essentially trying to build the production methods of five or ten years from now.
Another change, I think, relates to changes within industry itself.
Large-scale industrial production used to solve the question of availability.
Did you have a car to drive? Clothes to wear? A phone to use?
But people’s lives have now changed.
Of course I have a car, but which car do I want to drive? Of course I have a phone, but which specifications do I want? If I buy a gift, can it be different from everyone else’s?
Actual production demand has therefore gradually shifted from the large-volume, repetitive production of the 1980s, 1990s and even the 2000s toward increasing numbers of small-batch, customized requirements.
Many component orders may contain only 100–200 units, or 300–500.
In those circumstances, a conventional fixed production line may no longer be suitable.
You cannot build a new line for 200 products.
If dexterous manipulation and AI enable one line to do the work previously done by five and accept ten different kinds of orders, the calculation is no longer “how many workers does this robot replace?”
You should calculate how much additional business you can take on because you have the system.
At least in the short term, that is how I think the economics should be calculated.
As for standards, I think the entire embodied-intelligence industry lacks them.
First, it lacks data.
Everyone knows that data is among the scarcest and most expensive resources in embodied intelligence today.
You can wear equipment to collect first-person data, use two- or three-fingered devices to simulate a robotic hand, or use the most expensive approach: teleoperating real hardware, combining an arm and a hand to reproduce the actual equipment completely.
There is not enough of this data. It also has a very troublesome problem: a lack of interoperability across embodiments.
When our hand and a competitor’s hand each collect their own data and map it through real-to-sim-to-real, the results may not be directly reusable between them at all.
Even the hardware interfaces differ today.
The industry is therefore nowhere near its “USB-C moment.”
I do not think future standards will necessarily begin with a national standard that everyone then follows.
Look at many interfaces and protocols today, including Skill and MCP. They essentially emerge from the bottom up.
It is not that someone first decrees, “Everyone will use this from now on.” Rather, more people find something useful, start using it, and gradually develop an industry consensus.
Embodied intelligence has not yet matured to that stage.
So what is missing?
Data, cross-embodiment interoperability and interface standards. Essentially all of them.
Li Jufan: Mr. Zhang just mentioned incompatibility across embodiments. There is a question I have been thinking about recently that I would like to raise spontaneously.
Large language models eventually produced companies such as OpenAI and Anthropic.
What will happen with robots?
Will each robot manufacturer build its own model, or will a foundation model eventually emerge that works across different manufacturers’ robots?
Zhang Long: That is an excellent question.
We have also tried developing world models and robotics models ourselves.
But frankly, having worked at Baidu, I know how the supposed combination of big data, massive compute and large parameter counts is built up.
Over the past few years, spending on a foundation model has risen from tens of millions to hundreds of millions, then billions.
An embodied foundation model will require more investment in the future, not less.
Given the scale of today’s embodied-intelligence companies, if every business funds its own foundation model, to put it bluntly, everyone may burn through money together and fail together.
I even think the companies that eventually produce dynamic world models may be today’s large-model companies.
They may not necessarily continue to rely on Transformers; new approaches could emerge.
But their accumulated expertise in compute, algorithms and training large-parameter models is difficult for most embodied-intelligence startups to match.
An embodied company still has to manufacture hardware and develop its core business. Asking it also to spend tens of billions on a foundation model puts it at a completely different starting line.
Our assessment is therefore that a true foundation model may not ultimately emerge from this group of embodied-intelligence companies funding it themselves.
It may instead come from the very large technology companies applying their established methods in collaboration with us, using our data, models and robotics experience for training.
We would then fine-tune on a better world model.
Li Jufan: Understood. There is indeed no answer yet. I have been discussing this with founders recently, so your point about different embodiments prompted me to ask it here.
Zhang Long: Right. It is neither very far away nor very close. People have different perspectives today, so their assessments will naturally differ.
Kang Zhengzhong: Another question for Mr. Li. Yard robots are consumer products, and many startups choose crowdfunding as their first step. But a large crowdfunding total does not mean the product will become a hit after actual delivery. Those are different things.
Kang Zhengzhong: Yard robots are also highly seasonal products. Many users may need a full season of use before they can genuinely evaluate them. Returns, exchanges, after-sales support and consumables all become costs.
Kang Zhengzhong: For a robotics startup, then, is it harder to raise an impressive crowdfunding sum or to complete its first real volume delivery? And how do you determine whether this is a sustainable business rather than a one-off crowdfunding campaign with nothing afterward?
Li Jufan: Both are difficult, but in entirely different dimensions.
Crowdfunding is more like a marketing activity.
My product has not entered mass production, and I have zero accumulated brand equity. I therefore need to explain its definition, selling points and advantages three months or even half a year in advance, finding users willing to try something new and pay ahead of time.
Those users also often have strong opinion-leader characteristics and considerable ability to spread the word.
Essentially, it is advance marketing.
Take the opposite extreme: why do major brands not need to do this?
Apple or Huawei can hold a launch event and go from announcement to sales in perhaps a week. The brand can attract huge amounts of traffic and attention simply by doing something.
A startup does not have that.
Crowdfunding is therefore about slowly building a fire from nothing, rather than triggering an immediate explosion.
From a startup’s perspective, crowdfunding has another very important function: validating product-market fit.
At least establish whether anyone is willing to pay for the concept.
But entering mass production after crowdfunding presents a completely different level of difficulty.
A startup’s team is often not yet stable or fully staffed, and the supply chain has only just been assembled. Structural components, electronics and hardware design may all be new, with no mature production process to copy directly.
You will therefore encounter problems all the way from engineering prototypes through trial production to mass production.
One thing is especially important at that point: first, you must be certain your product definition is right.
Some companies finish crowdfunding while it is still questionable whether their product definition holds up.
You may have satisfied the needs of only a very small group.
If you are confident in the product definition and have thought through its differentiated positioning and user needs, much of the mass-production phase is not particularly glamorous.
What matters most is that the team pulls together.
I have seen early defect rates above 50%. That is not unusual.
You have to work through it step by step, solving each problem in turn.
The first crowdfunding customers actually offer an advantage: they are usually willing to accept a less stable product and provide feedback.
At that point, I think team cohesion and the founder’s strength of resolve are particularly important.
Of course, there are also many tactical things you can do.
For example, during small-batch production, supply some key opinion consumers first. They are early users who can later help create content and may gradually develop into key opinion leaders.
But those are still tactical measures.
What really determines whether you get a result is whether the team persists in making the product work.
Crowdfunding is therefore advance marketing, product-market-fit validation and product-concept validation.
What is mass production?
It is the moment when what is in your presentation finally has to become a real product.
What is the greatest fear then?
The presentation envisions a young woman, but the mass-produced result becomes a middle-aged auntie. That may be the end of it.
After mass production, the first priority for a first-generation consumer-electronics or smart-hardware product is not especially glamorous either: do after-sales support well.
Technical support, customer-service resources, a knowledge base and FAQs all have to be in place beforehand.
Do not wait for problems to appear before building after-sales support. Work backward from the product’s arrival: how far in advance do I need those resources ready?
Supplies must move before the army does.
Believe me, a first-generation product will have all kinds of bugs.
Many problems stubbornly refuse to show up in a test laboratory, then appear as soon as the product reaches real users.
In the most extreme case, if a highly influential user starts publicly criticizing your product, you may genuinely need to buy a plane ticket, fly over and solve the problem on site.
If you resolve it well, that person may instead become a positive opinion leader for you.
I have incubated new categories within large companies and also founded businesses myself, going from zero to one several times.
My strongest impression along the way is that there are not so many especially glamorous things involved.
Everyone knows the methodologies.
Huawei has IPD; some companies use EVT, DVT and PVT. After more than five years in the field, people generally know what should happen at each stage of product development.
What really matters is whether you thoroughly execute the necessary actions at each critical point.
Have the resources been prepared in advance?
Has every after-sales action been followed through to completion beforehand?
Much of the work before and after mass production is dirty work.
After-sales support is even more so.
In my most recent venture, I oversaw marketing, operations, sales and after-sales support—the entire customer-facing operation.
When products first arrived, what did we do whenever a problem appeared in the community?
Simple: if there were not enough people, everyone became customer service.
We made a roster.
Someone had to be on duty every day from 8 a.m. until midnight.
An influential user posting a negative review in your Facebook Group and getting 30–50 likes could directly affect conversion the next day.
Why?
Because the early user pool is only so large.
In consumer businesses, the first wave is called early adopters. Those few thousand people from crowdfunding are your “angels.”
Serve your angels well and you will be rewarded. Serve them badly and there really will be consequences.
It is very practical. That is how it works.
Kang Zhengzhong: How do you determine whether it is a sustainable business?
Li Jufan: Business-to-business and consumer businesses have some similarities in their underlying logic, as well as differences.
In B2B, you identify a lead, find an opportunity, close the first deal, then try to turn it into a success story—a reference case—and replicate it with a second and third customer.
What about consumer businesses?
You first reach a group of early adopters, then determine whether you can extend that group to a broader population.
The first step is therefore to ensure that what you presented really gets delivered and that after-sales support holds up.
The second is to understand who the first users actually are, then look for a second, broader group.
For example, conduct post-launch research.
On the product side, look at net promoter score: are users actually satisfied? Is the positioning right? Do users really care about what you consider the most important selling point?
On the marketing side, ask who the first users are. How old are they? Where do they live? What experiences and interests do they share?
Only after learning those things do you know where to find your target audience.
Should I look for opinion leaders to review the product on YouTube or TikTok?
Are my users Generation Z or Generation Y?
Should I use mainstream platforms, specialist media or niche communities?
Find what the first wave of users has in common, then expand to a second wave. That is a common approach.
Once you have built attention online, negotiating with offline channels becomes much easier.
Kang Zhengzhong: Another question for Mr. Dou. Automation in medicine and life sciences often involves many departments: the actual equipment users, equipment providers and those controlling the budget.
Kang Zhengzhong: When automating a laboratory workflow, which metric ultimately has the greatest influence on an order? If you had to rank throughput, accuracy and traceability, how would you order them?
Dou Haotong: Let me first step back to Mr. Zhang’s point about the “GPT moment.”
Personally, I think physical AI’s GPT moment may not arrive suddenly on one particular day. It may come gradually.
Why?
For the past six months, I have consistently taken Pony.ai’s robotaxis. They are sometimes slow and not necessarily that convenient, but I keep using them for commuting or meeting friends.
They have given me a strong impression:
Once you genuinely begin accepting the technology, it only needs to increase its vehicle installations and coverage incrementally to become part of your life.
Ordinary people may not even realize that it is gradually entering society.
Of course, a large-model company could suddenly produce an especially intelligent model one day, creating a real GPT moment in that way too.
Returning to the question:
For our company’s customers, product stability comes first.
If we integrate more AI in the future, its ability to control and operate the entire laboratory-automation workflow will also matter.
But from the end user’s perspective, one core problem many laboratory machines have long failed to solve properly is traceability.
Life sciences places great importance on making every experimental process traceable.
If you have followed research-integrity issues in life sciences for a long time, you will understand why provenance matters so much.
I therefore think traceability will be a very important part of laboratory AI in the future.
But how exactly will it be achieved?
Frankly, we do not yet have a final answer either.
Kang Zhengzhong: I would like all four of you to answer the final question. Whether for businesses or consumers, robotics products inevitably contain some nonstandard elements. But in volume delivery, you cannot customize everything anew for every customer and every region.
Kang Zhengzhong: How, then, do you gradually turn those nonstandard problems into standardized products as you move from a demonstration installation to volume orders?
Hu Xiaoping: In our experience, the key is to find a shared technical problem and use the same core technology to address applications across multiple industries.
Once that common technical problem is solved, the most important part is already standardized.
Adding other equipment or peripheral adaptations then allows it to be replicated fairly effectively as a relatively standardized product.
Kang Zhengzhong: In other words, first identify the most important shared problem and standardize the core technology. Localize or adapt the less central parts to the application afterward.
Zhang Long: I think the key is still one phrase: limited intelligence within a limited space.
A hand might theoretically do 1,000 things.
But in a particular factory, workstation or production stage, it may only need to perform ten of them frequently and well.
The remaining 990 functions might be used only twice a year.
Yet those 990 functions may require ten or twenty times the cost.
In industrial settings where ROI is scrutinized to the limit, that does not make economic sense.
Cut what needs to be cut. Remove those unnecessary capabilities.
Focus on the tasks that are genuinely most frequent and most valuable.
Industrial applications often require precisely that compromise.
Kang Zhengzhong: Subtract cost, weight and complexity.
Zhang Long: Exactly. Subtract.
Li Jufan: To me, that sounds like finding the “greatest common denominator.”
On the consumer side, once scale builds, it becomes more like the smart-hardware and consumer-electronics playbook.
Phones ultimately form different price tiers with different amounts of RAM and storage and different screen sizes.
Robotic mowers and pool cleaners are the same.
There are large and small lawns, steep and gentle slopes.
If a lawn is small and relatively flat, mowing it takes the user less effort, and hiring someone is not very expensive either. The product they need can naturally be simpler.
If the lawn is very large with many slopes, doing it takes more time and hiring someone costs more. The robot then needs greater cutting and climbing capabilities.
Ultimately, therefore, core capabilities such as cutting and slope handling are standardized, with different price tiers and product forms created for different applications.
Essentially, it is a portfolio of products at multiple price points based on user scenarios and purchasing power.
Dou Haotong: Finding the greatest common denominator is relatively straightforward for us.
Our customers are mostly researchers. Using our machines naturally generates substantial experimental data and research findings.
Around 15,000–16,000 biological laboratories worldwide now use our equipment, including many leading life-sciences research laboratories.
After those laboratories use our equipment for different experiments, we can examine large amounts of data and real usage scenarios to see what everyone is actually doing.
Once the data accumulates, we look for the greatest common denominator within it.
That is one way we gradually turn nonstandard requirements into standardized products.
Kang Zhengzhong: The four fields have ultimately given somewhat different answers. Industrial robotics looks for shared core technology across industries. Dexterous manipulation deliberately subtracts capabilities within a defined application. Consumer service robots standardize core capabilities, then use different SKUs to cover different scenarios and price bands. Life sciences works backward from large amounts of real experimental data to identify the greatest common denominator.
Kang Zhengzhong: That is all we have time for today.
This is a complete English translation of Unique Research’s Chinese article published on September 7, 2026, including its editorial introduction and full panel discussion. The original is available at https://mp.weixin.qq.com/s?__biz=MzU5Mjg5MjQ5Ng==&mid=2247522495&idx=1&sn=479869c97e195f43842a6aa2c2c44205 . The organizer’s agenda at https://luma.com/szopenday2026 identifies the panel participants.
Commercial figures, equipment lifetimes, laboratory practices, market estimates and forecasts are the original article’s or named speakers’ statements, not independently audited findings. Relative dates refer to the original conversation. Where the Chinese source does not name a currency, the translation does not supply one.
DexRobot and Lymow are the English brand names used for 灵巧智能 and 来牟科技. “Hechuang Shengwu” transliterates 合创生物; Opentrons identifies its Shenzhen entity as an Opentrons subsidiary. The speaker’s financing and customer references are retained as stated and should not be read as separately audited subsidiary-level figures. Li Jufan is listed as a former Lymow partner and CMO in the source guest list; his first-person account is retained without asserting a current role.
The original contains an Alzheimer’s analogy and an age-and-gender comparison when discussing product performance. These are retained as the speakers’ rhetoric, not medical explanations or editorial judgments about those groups.