---
title: "The Cooler It Looks, the More Likely It's a Fake Need: Four Overseas Companies Tear Back the Curtain on Physical AI Commercialization"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-10T18:21:07+00:00"
canonical: "https://ffcap.cn/en/research/the-cooler-it-looks-the-more-likely"
source: "https://uniqueresearch.substack.com/p/the-cooler-it-looks-the-more-likely"
language: "en"
---

# The Cooler It Looks, the More Likely It's a Fake Need: Four Overseas Companies Tear Back the Curtain on Physical AI Commercialization

[![](https://substackcdn.com/image/fetch/$s_!VZEb!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb05d815c-788b-4d47-8633-132de98285ca_2048x1152.jpeg)](https://substackcdn.com/image/fetch/$s_!VZEb!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb05d815c-788b-4d47-8633-132de98285ca_2048x1152.jpeg)

The cooler something looks, the more likely it's a fake need. The real signals hide in unsexy words: delivery, penetration rate, repurchase.

Three months, one million dollars, seven hundred overseas users paying real money.

Yisi Robotics' AI tennis robot posted this scorecard on Kickstarter — a result that would count as impressive for any hardware startup.

But at the roundtable during this Shenzhen Physical AI Summit, when Carina, Yisi's Marketing Director, talked about this experience, the focus wasn't on that million dollars at all.

"Basically every pitfall a hardware startup should step into, we stepped into all of them." Supply chain ramp-up, defect rates, logistics — all of it.

This roundtable was moderated by Wang Lu from TMTPost Going Global Reference. Four people on stage, exactly four ways to survive in physical AI going overseas: Hu Jiayi from Star Dynamic Era, putting humanoid robots into logistics warehouses; Zhu Yuanyuan from Orca Robotics, selling robot lawn mowers onto European and American lawns; Carina from Yisi, mass-producing a tennis training robot that moves around the court; and Wu Jiabing from Wondershare, doing software going overseas for over twenty years, with 90% of customers overseas.

Their businesses are very different, but by the end, every conversation converged on the same thing: hype and business are two different matters.

The cooler something looks, the more likely it's a fake need. The real signals hide in unsexy words: delivery, penetration rate, repurchase.

Let's start with the business with the most money.

Since mid-2025, Star Dynamic Era has been doing commercial validation with top logistics platforms like SF Express and China Post, going head-to-head with other solutions in the industry. Single-customer orders in the tens of millions have already emerged.

Why are customers willing to place such large orders? Hu Jiayi's answer contains not a single high-sounding word — just four metrics.

\*\*First, cost.\*\* Their main model, the M7, is modified from the full-size bipedal humanoid robot L7: upper body with dual arms, lower body replaced with a wheeled column. For robots to enter logistics, cost is unavoidable. As delivery scales from thousands to tens of thousands of units, cost must come down rapidly.

\*\*Second, efficiency.\*\* Logistics scenarios don't care about performance art. "It doesn't matter if you do a beautiful move — it just needs to keep grabbing packages nonstop all day." Customers only ask one thing: how many pieces can you process per unit of time?

\*\*Third, accuracy — and not laboratory accuracy.\*\* What does a real logistics site look like? Irregularly shaped items, damaged items, non-standard packaging, stained surfaces, wet surfaces, inconsistent label orientation, changing light conditions, production lines running 24 hours nonstop. The robot can't just pick things up — it also needs to flip, recognize, scan, and place labels face-up.

\*\*Fourth, scaled delivery.\*\* Building one unit in a lab is very different from delivering hundreds or thousands at once.

When the industry first started commercial validation, the target was set at 85% of human efficiency — reaching that number already meant you could replace some workers. Then everyone kept pushing toward 90%, 95%, 98%.

As for how the price of those tens-of-millions in large-customer orders is negotiated, Hu Jiayi gave an honest answer: the actual transaction price is a trade secret. But when large customers initially take only dozens or a hundred units, they typically don't negotiate based on the cost logic of those dozens of units — they negotiate based on the price of future scaled delivery.

"Customers are paying for the future cost you're promising. You have to actually deliver the volume."

Put simply, customers are betting not on the cost of the current dozens of units, but on your cost after future scaling. This is the coldest and most honest logic in logistics scenarios.

Why is this business worth doing? Hu Jiayi has visited large logistics centers himself: packages surge in like a flood, piling up into mountains even on ordinary workdays. Humans get eye strain, fatigue, emotions, and occupational diseases — while logistics is precisely an industry with a huge market and extremely thin margins, naturally sensitive to efficiency. The value of machines in such places needs no imagination to justify.

Yard robots have seen brands pile in over the past few years, with technical routes argued into a tangle — vision, RTK, LiDAR, each with its supporters. Zhu Yuanyuan didn't agonize over which route was more advanced; he talked about how Orca made its choice back then.

Three or four years ago, smart robot lawn mowers were far less mature than today, and it was impossible to see which route would become mainstream. Orca asked only one question at the time: this product has been sold for so many years — why hasn't it truly entered every household?

The answer was stuck in the shovel in the user's hand.

When early robots were brought home, the user's first task wasn't mowing the lawn — it was digging up the ground in their own yard and burying a boundary wire. The robot relied on this wire to know where their lawn ended, so it wouldn't charge into the neighbor's yard. Imagine the scene: you spend a lot of money on a smart product, and the unboxing ceremony is first breaking ground in your yard.

So the first direction for the new generation of robot lawn mowers was boundary-free. Orca started betting on the 3D LiDAR route in 2023 — no wire burying, no extra base stations, no reliance on external networks, and no fear of positioning being blocked by trees or houses.

The only drawback was cost. Three years ago, a single LiDAR cost 3,000-4,000 or 4,000-5,000 yuan. Later, costs dropped all the way to the hundred-yuan level. A route that was very non-mainstream three years ago is now slowly becoming mainstream.

Zhu Yuanyuan's summary is worth copying down for anyone doing hardware: don't first ask which technology is most advanced — first ask what the original pain point of this category was, then reverse-engineer the technical route.

As for the price war everyone in China is shouting about, he gave a very different explanation: the previous generation of wire-requiring robots holds massive inventory overseas. As the new generation of boundary-free products rises, if old products don't cut prices to clear inventory now, they'll be even harder to sell later. Some price wars are essentially holding funerals for the previous generation of products — just like the wave of price cuts for fuel cars when new energy vehicle penetration rose.

And as for the judgment that this is a "red ocean," he directly pushed back with data: many people in China think robot lawn mowers have become a red ocean, but if you look at North America and Australia, real penetration is still in single digits — a huge number of users have never even seen this product.

"High market volume doesn't mean the market is already done."

The moderator asked: Will you make humanoid robots in the future, directly going on court with rackets to play against people? After all, videos of humanoid robots playing tennis are really exciting and get a lot of traffic.

Carina said: "Mechanical form should serve function, not imitation."

Her reasons were all about the math. Today, the cost, stability, and battery life of humanoid robots can't support an ordinary consumer taking it to the court for an hour. A single motion capture setup costs hundreds of thousands of yuan — and that's not even counting the robot itself.

Then there was the fan analogy: "Even if every household has a humanoid robot in the future, you won't necessarily make it fan you just because it has two hands. Because air conditioners already work very well."

So Yisi chose a wheeled structure: cheaper, more stable, more suitable for this task. When legs truly improve performance on the tennis court and cost is low enough, they can consider it then — no rush.

Behind this judgment is their order of operations. The core pain point of tennis training is dependence on sparring partners and coaches, and people are all constrained by time. Traditional ball machines are one-way: the ball is sent to you and it's over. Their robot can see your ball, judge the landing point, move over and hit it back, completing near-real continuous rallies, and can gradually identify your weaknesses.

First break through the most fundamental thing, then add features on top. Not the other way around.

But after validating the need, the really hard days begin. Crowdfunding only proves that someone wants to buy — it doesn't prove you can deliver.

"Crowdfunding is like an engagement. Delivery is actually living together."

Traditional ball machines need just over fifty components. After adding movement, recognition, and interaction, Yisi's complete machine has over two hundred components — the complexity is on a completely different level. Once the robot actually starts moving, all kinds of bugs emerge that were never seen in the Demo stage. They can only be thrown into the real market for validation, then pulled back for iteration.

Even something as small as the ball bag tripped them up: should it be a trolley-style or a cup-style? In the end, users gave the answer: everyone is already used to carrying a ball bag to the court, so combine the two — it's both the robot's ball bag and the user's own ball bag.

After several months of iteration, hundreds of units have been shipped overseas one after another. And Carina's long-term judgment on this business: "Hardware determines the floor, software defines the ceiling."

Hardware delivery is only the first step. Behind it are training content, personalized training, and AI coaching. One day, when the system finds your backhand has a particularly high net-fault rate, the next time you open the app, it directly recommends a backhand-specific training program.

Wu Jiabing from Wondershare was the only person on stage not doing hardware, but his judgment framework is worth thinking about.

After making products for over twenty years, he increasingly cares about one question: what is a real need, and what is a fake need?

Some products look particularly cool — high traffic, even decent user numbers. But look closer and you'll find low payment and low usage frequency. He tends to judge these as fake needs.

Conversely, some needs are particularly small — like format conversion in software. It doesn't sound sexy at all. But when users truly need it, it solves that problem extremely accurately, and in the end becomes a stable cash flow for the product: stable user base, stable retention, stable willingness to pay.

"As long as it truly solves a problem, even the smallest need is a real need. If it just looks cool and the business model can't be explained clearly, it's very likely a fake need."

Apply this standard to the currently hottest AI comic dramas, and the picture immediately cools down. The market is noisy, but the ROI of paid content has already dropped to around 1.1 — basically rubbing against the breakeven line. The dividend of live-action short dramas was released over several years; AI comic dramas started reshuffling in less than half a year.

The opportunity is of course real too: production cycles compressed from months or one-to-two years to days or even hours; the cost of a live-action drama in the millions has been compressed by AI content to the thousands-of-yuan level; a large amount of overseas comic drama content is already produced by Chinese teams, and industrialization capability is spilling outward.

As for the methodology of going overseas, he paid tuition for three pitfalls.

\*\*The first pitfall is feature stacking.\*\* Users like one feature, so you add a second, a third — until the core feature users actually need can't be found anymore. Hardware is the same: today the robot can mow the lawn, so you want it to trim branches, sweep leaves, remove weeds — make it omnipotent, and in the end nothing works well.

\*\*The second pitfall is compliance.\*\* When software has problems, you can still update remotely — hardware can't. Goods all shipped to the US and Europe, and because one certification can't be sold, the loss is enormous. Hardware compliance must be front-loaded.

\*\*The third pitfall is mistaking localization for translation.\*\* Wondershare established a company in Japan back in 2008. A real Japanese market requires local employees, understanding local culture, participating in local events. No matter how authentic the language translation, it doesn't mean users are willing to pay.

His feeling from visiting Silicon Valley this year: a very vertical, very small need can also support a startup. Users will tell you very clearly: solve this problem for me, and I'll pay.

For the last question, the moderator asked everyone: in the industry, what is a real signal, and what is just noise?

Hu Jiayi was direct. What's truly certain is that the entire industry has reached consensus on the importance of embodied models, on the "brain" — but how exactly to do it, the technical routes are still diverging.

The noise is also obvious: "Right now, everyone on the street is talking about world models." Those making brains talk about it, those making bodies talk about it, those making dexterous hands, joints, electronic skin talk about it — even traditional precision manufacturing companies have started talking about it through investment and acquisition. New concepts are of course needed, but overemphasis easily becomes跟风 (following the trend). Embodied intelligence today确实 has overheating.

But concept aside, the real problem behind world models exists: high-value data from the real physical world is currently very scarce. The concept may be noise, but the data gap is a real signal.

Carina's noise list was shorter: Demo.

Many people see a Demo made and think the product is already mature and usable. But what robot companies really look at are other questions: Can this product be delivered? How many people are actually willing to pay? Of those who pay, how many will keep using it? Can this company maintain delivery and service long-term at a reasonable cost?

Demo only represents that the technology was implemented once. From product to business, every one of these questions stands in between.

Near the end of the roundtable, the topic turned to US FCC regulation of Chinese smart hardware. Zhu Yuanyuan told a case he recently saw, which I'll save for last.

A peer company performed very well on a crowdfunding platform, but because of policy issues sent a letter to users saying shipment might be temporarily unavailable. According to the script, what should follow is a wave of refunds and negative reviews.

Instead, in the comments section, users were teaching the company how to do business: "Don't cancel my order — I can go pick it up in Canada." "Ship it to me as disassembled parts, I'll assemble it myself."

This is probably the best footnote to the entire roundtable. Concepts become outdated, Demos become outdated, today's hot technical route may in three years be the previous generation that needs clearance. But a user willing to cross a border to pick up goods doesn't lie.

The need is real. The only question left is: make the goods, ship them out.

\*\*Wang Lu\*\*: Our discussion is about business models — the overall logic goes from robot demonstration, to real delivery, to finally running positive profits. The first question, combined with self-introductions: why do users need to buy your products? What is the point they're truly willing to pay for?

\*\*Hu Jiayi\*\*: I'm Hu Jiayi from Star Dynamic Era. Star Dynamic Era positions itself as a software-hardware integrated, full-stack self-developed general embodied intelligence company — from embodied brain, body, dexterous hands, all the way to scenario data and commercialization landing.

Why do customers buy our products? I think the core is whether the entire chain can truly close the loop. From brain, body, dexterous hands to scenario data, we ultimately need to achieve hand-eye coordination and truly put capabilities into scenarios for validation. Right now a core scenario is logistics — we've done quite a bit of commercial validation in the logistics industry.

\*\*Zhu Yuanyuan\*\*: I'm Zhu Yuanyuan from Orca Robotics. We mainly make overseas yard smart products, with robot lawn mowers as the core.

Why do users buy? Essentially because it solves a real need. Many European and American households are different from China — the lawn itself is the "face project" of the home. Every year, a lot of time, energy, and money goes into maintenance.

Robot lawn mowers at least directly solve one thing: handing over repetitive, heavy physical work like mowing to robots. What users save isn't just physical effort — it's more about time. They can use that time for living, for being with family.

\*\*Carina\*\*: I'm from Yisi Robotics. We make AI tennis robots that have achieved mass production. The selling point is actually very direct.

Traditional ball machines can only serve from a fixed position, but our robot can move and rally — it can judge where the ball lands, then move over and hit it back to the user, giving you an experience closer to a real human sparring partner.

We previously did a Kickstarter crowdfunding — nearly $1 million in three months, with about 700 overseas users validating this need with real money.

\*\*Wu Jiabing\*\*: I'm Wu Jiabing from Wondershare, mainly responsible for the company's strategy, investment, and M&A. We're different from the others — mainly software products, and 90% of our customers are overseas.

Why do users buy our products? We often don't solve particularly grand problems, but rather small pain points that users encounter frequently in daily work and life. When these small needs truly help users solve problems, they'll pay.

\*\*Wang Lu\*\*: Mr. Hu just mentioned that logistics is a very important scenario for Star Dynamic Era, and single-customer orders of relatively large amounts have already emerged. From the earliest testing to finally customers willing to do batch delivery, it's certainly not solved by just saying "the robot is smart." Why are customers willing to place such large orders? What metrics do they truly assess you on — efficiency, cost, or something else?

\*\*Hu Jiayi\*\*: Several aspects. Since mid-2025, we've been doing early commercial validation, PoC, with some top logistics platforms — including SF Express and China Post — and also done PKs with other solutions in the industry. Different companies take different technical paths.

Whether a logistics robot solution can truly land, I think at least four metrics matter.

\*\*First is cost.\*\* Our main robot right now is the M7, based on the full-size bipedal humanoid robot L7 — upper body is dual arms, lower body changed to a wheeled column. Later the company will have products fully designed for industrial scenarios.

As delivery scale goes from thousands further up to tens of thousands, we hope the entire cost can continue to drop rapidly. For robots to truly enter logistics, cost is definitely unavoidable.

\*\*Second is efficiency.\*\* Logistics scenarios are very realistic — customers definitely look at how many packages can be processed per unit time. It doesn't matter if you do a beautiful move; it just needs to keep grabbing packages nonstop all day.

The industry's efficiency improvement is very fast now, and peers already have very high per-hour processing data. Our own efficiency over these past few months is also continuously improving.

\*\*Third is accuracy.\*\* Logistics sites aren't as clean as in Demos. Various irregular packages, damaged packages, inconsistent label directions, changing ambient light — all of these will be encountered.

So your robot can't just "pick it up" — it also includes flipping, recognizing, scanning, placing labels face-up. Once truly entering the production line, accuracy is very critical.

\*\*Fourth is scaled delivery capability.\*\* Achieving lab effects with robots is very different from actually delivering hundreds or thousands at once. Currently we still need to accelerate in this area; several leading companies in the industry have already achieved relatively large delivery scales, and we're now moving toward higher volumes.

The value that SF Express, China Post, and other large industrial customers give us isn't just buying equipment. They're also very important scenario partners. You must enter the real production environment to know whether this thing actually works.

\*\*Wang Lu\*\*: From when you first started until now, how much has the robot's real cost actually dropped?

\*\*Hu Jiayi\*\*: The specific real transaction price for top customers is a trade secret, not convenient to discuss. But when large customers initially only take dozens or a hundred units, they generally don't negotiate completely based on the cost logic of those dozens or hundred units — usually they negotiate based on future long-term cooperation, scaled delivery. The price achievable after scaling is itself the direction the entire industry is working toward now.

\*\*Wang Lu\*\*: Mr. Zhu, yard robot brands have been very numerous these past few years, both domestic and international, with very fierce competition. There are also many technical routes — vision, RTK, LiDAR, etc. Why did Orca choose the current route back then?

\*\*Zhu Yuanyuan\*\*: We didn't start by "choosing a technical route for the sake of choosing a technical route." It should be looked at in reverse: what problems are users actually encountering? What technology can solve this problem best?

Three or four years ago, smart robot lawn mowers weren't as mature as today, and which technical route would become mainstream was completely unclear at the time. What we considered then was: why hasn't this product truly entered every household on a large scale after so many years?

In the end, we found the most fundamental problem was still user experience. After early robots were brought home, the user's first task might not be mowing the lawn — it was digging up the ground in their own yard and burying a boundary wire.

Why? Because the robot needed this wire to know where their lawn ended, so it wouldn't run to the neighbor's. This usage threshold was actually very high.

So later, the first direction for the new generation of robot lawn mowers was boundary-free.

And how to solve boundary-free? You can use RTK, or you can find a way to completely not rely on external positioning facilities, letting the robot achieve autonomous positioning through its own sensors — like cameras, or 3D LiDAR.

Orca started pushing the 3D LiDAR route around 2023. We judged at the time that this route would have better user experience in the long term.

No wire burying, no extra base stations, not so dependent on external networks, and positioning won't be affected by trees or houses blocking. The only problem is — it's expensive.

Three years ago, a LiDAR might still cost 3,000-4,000 or 4,000-5,000 yuan. But later, costs had dropped to the hundred-yuan level. Once costs came down, you find that a route that was very non-mainstream three years ago is now slowly becoming mainstream.

So I think technical routes ultimately must return to user experience. Don't first ask "which technology is most advanced" — first ask: what was the original most painful point of this category? Then reverse-engineer the technical route and product version.

\*\*Wang Lu\*\*: Carina, why did Yisi choose sports, especially the tennis scenario? Some people are now making humanoid robots play tennis — why did you choose the current product form?

\*\*Carina\*\*: We think tennis is actually very suitable for robot entry. Because tennis training has a particularly realistic problem: it's highly dependent on sparring partners and coaches. But whether sparring partners or coaches, they're all highly dependent on human time.

You have time today, the other party might not; the other party has time, you might not. So what we thought at the beginning wasn't "how can a robot replace a person" — but whether it can provide a stable, continuous, precise training output.

Traditional ball machines are one-way. They send the ball to you, and it's over. We more want to make it two-way interaction.

There's a robot across from you that can see your ball, judge the ball's position, then move over and hit it back, completing a near-real continuous rally with you. It can also gradually identify your weaknesses. That's the core problem we want to solve.

So we also first break through the most fundamental thing, then add features on top.

\*\*Wang Lu\*\*: Will you consider making humanoid robots in the future, directly going on court with rackets to play against people?

\*\*Carina\*\*: Mechanical form should serve function, not imitation. Videos of humanoid robots playing tennis are indeed very exciting, very cool.

But today, their cost, stability, and battery life are all still insufficient to support an ordinary consumer taking it to the court for an hour. Motion capture equipment alone might cost hundreds of thousands of yuan — not even counting the cost of the humanoid robot itself.

Even if every household has a humanoid robot in the future, you won't necessarily make it fan you just because it has two hands. Because air conditioners already work very well.

So we currently choose a wheeled structure, because it's cheaper, more stable, and more suitable for this task. In the future, if legs truly can significantly improve performance in the tennis scenario and cost is low enough, we'll of course consider it. But you can't make it humanoid just because "it looks like a person."

\*\*Wang Lu\*\*: Mr. Wu, Wondershare has been globalizing for over twenty years and has also done a lot of investment. From a product and investment perspective, many AI and embodied intelligence scenarios now may not have that large a user scale, but indeed some people are willing to pay. How do you view this kind of opportunity?

\*\*Wu Jiabing\*\*: After making products for over twenty years, we increasingly care about one question: what is a real need, and what is a fake need?

Some products look particularly cool. Traffic is also large, user numbers even quite a few. But if you truly look inside, you'll find payment is low and usage frequency is also low. We tend to judge this kind of need as a fake need.

Conversely, some needs are particularly small. Like format conversion in software. It doesn't sound sexy at all. But when users truly need it, it can solve that problem very accurately.

This kind of need, in the end, may instead become a stable cash flow source for a product. Stable user base, stable retention, stable willingness to pay.

So my own judgment is: as long as it truly solves a problem, even the smallest need may be a real need. Conversely, if the business model can't even be explained clearly, and it just looks particularly cool, then it's very likely a fake need.

\*\*Wang Lu\*\*: Can you, without naming names, talk about a fake need in embodied intelligence?

\*\*Wu Jiabing\*\*: Actually quite a few. If you put ten embodied intelligence products on stage, many times you can tell at a glance which ones are truly solving problems and which are more like flower stands. Everyone actually has a gut feeling.

\*\*Wang Lu\*\*: Wondershare is also doing new business layout now, investing in AI comic drama-related enterprises. How do you view the opportunities and challenges of this track?

\*\*Wu Jiabing\*\*: AI comic dramas have indeed been very hot this year. The market and users are both growing rapidly. But if you truly look at commercialization, you'll find that the dividend in this industry fades particularly fast.

Right now, the ROI of much paid content is already close to 1.1, basically already near the breakeven line. So the surface looks very lively, but underwater it's actually very fierce.

The dividend of live-action short dramas may have been released over several years. AI comic dramas may have started reshuffling in less than half a year.

But it also brings several very obvious opportunities.

\*\*The first is AI tools' improvement of production efficiency.\*\* Previously, a piece of content, from production to completion, might take months, even a year or two. Now the entire production cycle has been greatly compressed, already starting to be calculated in days, or even hours.

\*\*The second is a large drop in production cost.\*\* Previously, a live-action drama might need millions of yuan; now the production cost of AI content may have dropped to the thousands-of-yuan level — the core is computing power and production cost.

\*\*The third is going overseas.\*\* Right now, a large amount of overseas comic drama content is itself produced by Chinese teams. After tools mature, China's这套 content production industrialization capability will rapidly spill outward globally.

But the challenges are also very clear.

\*\*First is character consistency.\*\* Episode one and the final episode, the same character might look different — this is still a technical problem.

\*\*Second is localization.\*\* Many people used to think "going overseas" means translating Chinese into English, Japanese. That's not it.

We established a company in Japan back in 2008. To truly do the Japanese market, besides translating the language, you also need local employees, understanding local culture, participating in local events. True going overseas isn't language going overseas — it's culture going overseas.

\*\*Third is channel dependence.\*\* Many Chinese enterprises used to have a misconception: as long as I find an overseas agent, they'll solve all problems for me — sales, user needs, product feedback, localization. Later we found that's根本 not the case.

If you don't participate in local operations yourself, don't directly contact users, don't understand user pain points yourself, in the end you'll likely make many so-called "needs" that aren't real needs.

\*\*Wang Lu\*\*: Carina, your Kickstarter crowdfunding has been close to $1 million. Users are willing to buy, so the need can be considered validated. But for many hardware companies, the truly most painful stage is often from crowdfunding to delivery. What pitfalls have you stepped into?

\*\*Carina\*\*: We actually don't simply understand crowdfunding as marketing. For us, its more important role is doing function and performance validation.

Over 700 Kickstarter users have already told us with money that they truly need a long-term, efficient tennis training partner, not just a fixed ball machine. But after crowdfunding, we indeed encountered delivery problems.

The good news is that after several months of iteration and upgrades, hundreds of units have now been shipped overseas one after another. The bad news is that basically every pitfall a hardware startup should step into, we stepped into all of them. Supply chain ramp-up, defect rates, logistics — all of these were encountered.

The company's focus has also shifted from the initial: "How to make users willing to buy?" to: "How to deliver stably?" "How to truly get the product into overseas users' hands?" "How to make users keep using it after receiving it?"

So between Demo and mass production, there must be such a stage. It can't be avoided.

\*\*Wang Lu\*\*: Why do these problems occur? Is it because new products don't have mature supply chains, components aren't standardized enough, or other reasons?

\*\*Carina\*\*: Traditional ball machines might need just over fifty components to complete basic serving. But after adding movement, recognition, and interaction, our complete machine might need over two hundred components. The complexity is completely not on the same level.

So the problem isn't just "whether the supply chain is mature." It also includes all kinds of bugs that emerge after the robot actually starts moving — bugs that were never seen at the Demo stage. These things can only be validated through the real market, then brought back for iteration.

For example, a very specific problem: should our ball bag be trolley-style, or cup-style?

In the end, through user validation we found that everyone is already used to carrying a ball bag to the court. So we later combined the two needs, making it both the robot's ball bag and usable as the user's own ball bag.

This kind of product needs to be continuously put in front of real users for validation. We're also very grateful to the first batch of crowdfunding users, willing to wait for us for several months, perfecting the product little by little.

\*\*Wang Lu\*\*: A big difference between AI hardware and traditional hardware is that software continuously provides services. So the business model may no longer be a one-time sale. Does Yisi have considerations for software subscriptions or other continuous charging later?

\*\*Carina\*\*: After traditional ball machines are delivered, it's basically over. But we care more about this person's subsequent training habits and training effects.

The machine should increasingly understand you in the future. For example, what level are you? How's your forehand? How's your backhand? Do you train every day? What should you practice in the next stage?

Suppose after you finish playing today, the system finds your backhand has a particularly high net-fault rate, many errors. The next time you open the app, it might directly recommend a backhand-specific training program.

So from the entire product logic, hardware delivery is only the first step. Behind it are power-on usage, training content, personalized training, and ultimately AI coaching.

We're now also continuously upgrading the training system, AI coaching, and coach-related app functions. I have an understanding myself: hardware determines the floor of this product, software defines its ceiling.

\*\*Wang Lu\*\*: Mr. Zhu, will yard robots also explore continuous payment?

\*\*Zhu Yuanyuan\*\*: Yes. First, robot lawn mowers themselves must rely on the app. Users open the app every day to check the robot's working status and maintenance results. This entry point naturally exists.

Currently we already have some continuous payment scenarios. For example, the standard 4G data plan — after it's used up, users can directly renew in the app.

Second, we've made mobile video monitoring. It's equivalent to a moving camera on the lawn. When users go on vacation, not at home, they still want to see what state their lawn is in — is it green or yellow, any abnormalities? This function needs extra data, so it can also form payment.

In the future, including accessories, consumables, value-added services, all may form subsequent revenue.

\*\*Wang Lu\*\*: There are all kinds of voices in the yard robot market now. Some think the growth inflection point hasn't arrived, some are already fighting price wars, and some teams have exited. Standing inside the industry, what's the real situation you see?

\*\*Zhu Yuanyuan\*\*: My feeling is that European and American consumers still have a large amount of deep needs unmet. Many products on the market today can probably only be said to "usable."

But there's still a long development process from truly being easy to use, practical, stable, and reliable. It's a bit like robot vacuums ten years ago.

The earliest robot vacuums didn't even have mopping or self-cleaning. They just finally could auto-plan and auto-map. At that time, at least they changed from "unusable" to "usable." Today's smart robot lawn mowers are actually relatively close to this stage.

As for price wars, I think part of it is caused by the alternation of new and old products. The previous generation of wire-requiring robot lawn mowers still holds massive inventory overseas.

After the new generation of boundary-free robots rises, if old products don't cut prices to clear inventory now, they'll be even harder to sell later. So some price wars are essentially clearing the previous generation of products. A bit like when new energy vehicles started rapidly increasing penetration, fuel cars also experienced similar changes.

\*\*Wang Lu\*\*: As a Chinese startup brand, what's the real bottleneck for commercial growth in the European and American markets?

\*\*Zhu Yuanyuan\*\*: Definitely exists, and the bottleneck isn't small. First, we're a startup, and a Chinese brand, with no long-term brand accumulation in the European and American markets in the past.

Starting from zero, making dealers and consumers believe you takes time. Besides product strength, you also need to leave enough channel margin for partners willing to promote you. So we're still in the exploration and expansion stage.

\*\*Wang Lu\*\*: Mr. Hu, you just mentioned that logistics has achieved relatively large-scale commercialization. What's the hardest hurdle in this scenario?

\*\*Hu Jiayi\*\*: If ranked, I think the first is efficiency, the second is accuracy. But this "accuracy" in logistics scenarios isn't laboratory accuracy.

A typical workstation we truly do is called small-piece package supply. This workstation has a very large workload, a large number of workers, and basically runs 24 hours nonstop.

What do real logistics packages look like? Irregularly shaped items, damaged items, non-standard packaging, stained surfaces, wet surfaces, plus on-site light and dark changes, noise, production lines running fast. All of these situations exist. This greatly tests the generalization capability of embodied models.

In terms of efficiency, when the industry first did commercial validation, the target might first be set at 85% of human efficiency. Reaching 85% actually already means you can replace some workers. Later everyone continues pushing toward 90%, 95%, 98%. The ultimate goal is of course close to 100%. But different companies have different testing standards, so numbers can't be simply compared horizontally.

Besides efficiency and accuracy, the third particularly important thing is: can the model adapt to the complex working conditions of real production lines.

If these three problems are truly solved, I believe some workers in logistics small-piece package supply will be replaced by robots relatively quickly. And as delivery volume ramps up, hardware costs, model routes, bodies, dexterous hands, joints — these things will also gradually start to standardize.

Once industry standards begin to form, the speed of robots entering logistics production lines will continue to accelerate. Logistics itself is a very large market, but margins are particularly thin. So enterprises naturally value efficiency very highly.

I've visited large logistics center sites myself. It's truly packages surging in like a flood, piling up into mountains even on ordinary workdays. Humans easily get eye strain, fatigue, plus emotions and occupational diseases. In this scenario, machines have very clear value.

\*\*Wang Lu\*\*: Let's discuss another globalization issue. The US is increasingly strictly regulating Chinese robots and smart hardware with software and communication capabilities. What impact does this change have now on several hardware-going-overseas enterprises? What does it mean in the long term?

\*\*Hu Jiayi\*\*: Currently the direct impact on Star Dynamic Era's body business isn't particularly large. In our overseas revenue, the proportion of bodies itself isn't that high — overseas is more dexterous hands, plus some model and technology licensing. Customers like universities and research institutions are relatively more.

So in the short term, business outside of bodies can still continue. But if you look at an earlier batch of Chinese robot companies — like cleaning, delivery, hotel robots — many of these enterprises already have a very high proportion of overseas revenue. This batch of companies truly survived in the cruel global market competition. They actually have reference value for later Chinese robot enterprises.

\*\*Zhu Yuanyuan\*\*: This policy will have a more obvious impact on smart hardware going overseas, especially categories like yard robots. Because the US itself doesn't have a particularly mature similar industry, and this year happens to be a year when smart robot lawn mowers are rapidly increasing penetration in the US.

From industry data, the US market grew very fast in the first half of this year. So new regulation appearing at this time point will definitely have an impact on growth expectations for the next few years — especially for enterprises that are launching new products in the US and still in the certification process.

But from another angle, real demand exists. This demand can't disappear with just a ban.

I recently saw a particularly interesting case. A peer company performed very well on a crowdfunding platform, but because of policy issues sent a letter to users saying shipment might be temporarily unavailable. As a result, in the comments section many users weren't asking for refunds — they were teaching the company: "Don't cancel my order, I can go pick it up in Canada." Some even said: "Ship it to me as disassembled parts, I'll assemble it myself." Users were反过来 helping the manufacturer figure out ways. This at least shows the demand is real.

\*\*Carina\*\*: In the short term, the impact on tennis robots is still relatively limited, because they're temporarily not in the core restriction scope. But it gives us a very clear signal: supply chain globalization is becoming increasingly important.

A Chinese hardware company wanting to do business overseas long-term can't just solve "whether the product can be sold." You also need international compliance, supply chain management, and local service capability.

So for Yisi, we hope to gradually evolve from the earliest relatively simple "channel going overseas" to brand, service, supply chain, and compliance going overseas together. This will become a variable that must be considered in the long-term business model.

\*\*Wang Lu\*\*: Mr. Wu, Wondershare has been going overseas for over twenty years. What experience is most worth referencing for today's AI hardware companies?

\*\*Wu Jiabing\*\*: When we earliest started going overseas, we mainly ate the traffic dividend. Later we continuously expanded products and features, and today it's become AI plus product innovation. In between, we stepped into特别 many pitfalls.

\*\*The first pitfall is feature stacking.\*\* At the very beginning, a product might have only one feature. Users thought it was good, so we continued adding a second, third, fourth. Until the end, you find that the core feature users truly need can't be found anymore.

Hardware is the same. Today the robot can mow the lawn, you'll start thinking whether it can trim branches, sweep leaves, remove weeds — preferably it can do everything. Doing and doing it becomes "omnipotent." But truly making every single thing work well is very hard.

Especially European and American users have different consumption habits from Chinese users. This year when I went to Silicon Valley, the feeling was particularly obvious. A very vertical, very small need can also support a startup.

Users will tell you very clearly: "Solve this problem for me, and I'll pay." So after Chinese products go overseas, they should instead make core features deeper, into true hard power.

\*\*The second pitfall is compliance.\*\* FCC is just one type. When software encounters compliance problems, it's relatively better because it can be updated remotely. Hardware is completely different.

If your goods have all been shipped to the US and Europe, and in the end because one certification or compliance issue they can't be sold, the loss is enormous. So hardware compliance must be front-loaded. Can't wait until the goods are all out, then passively make up for it.

\*\*The third is localization.\*\* Localization is absolutely not translating language. You go to the Middle East, Europe and America, Japan — different markets have different cultures, religions, color preferences, even many very detailed expressions are different.

No matter how authentic the language translation, it doesn't mean users will necessarily be willing to buy. True globalization requires making these details deep enough.

\*\*Wang Lu\*\*: Last question. The industries all four of you are in are now very hot, and also the areas most likely to appear in news sections. If you all judge, what in the industry recently is a business signal truly worth paying attention to? What is just noise?

\*\*Hu Jiayi\*\*: What's truly certain is that the entire industry has reached relatively clear consensus on the importance of embodied models, the embodied "brain." But how exactly to do it, the technical architecture and routes are still very diverse, still diverging.

What's the noise? I think a very obvious one is "world models." Right now, everyone on the street is talking about world models.

Those making brains talk about it, those making bodies talk about it, those making dexterous hands, joints, electronic skin, sensors also talk about it — even traditional precision manufacturing companies, through investment, acquisition, internal incubation, have all started talking about it. New concepts are of course needed. But after overemphasis, it easily becomes跟风.

Embodied intelligence today确实 has an overheating problem. Everyone swarms in, much like many industries when they first started over a decade ago. So new things still need critical thinking.

But on the other hand, the core problem behind world models确实 exists. Especially high-value data from the real physical world is currently very scarce. In the future, truly training usable models will require a very large amount of data. So the concept itself may have noise, but the real-world high-quality data gap is a very real signal.

\*\*Zhu Yuanyuan\*\*: Our industry is quite interesting. Many people in China now already think smart robot lawn mowers are a "red ocean."

But if you look overseas, especially North America and Australia, real penetration is actually still very low — even single digits. Many users have根本 never seen, nor truly used this product.

So we根本 don't know yet, when this product truly meets the needs of most households, what the final penetration rate will reach. But one thing I'm relatively certain about: it definitely won't stay at such a low level forever.

So high market volume doesn't mean the market is already done.

\*\*Carina\*\*: For tennis robots, a very obvious signal now is: robots that used to stay in Demos are gradually entering real scenarios.

Whether it's specialized robots, or the humanoid robots playing tennis that everyone has seen recently, this at least shows that robots have started走出 the laboratory, doing real interactions with real people.

But the biggest noise here also precisely comes from Demo. Many people see a Demo made and think this product is already mature, already usable.

But for robot companies, what we truly look at isn't how gorgeous the Demo moves are. What we truly look at is: Can this product be delivered? How many people are truly willing to pay? After users pay, how many will keep using it? And can this company, at a reasonable cost, maintain delivery and subsequent services long-term?

---

Original publication: https://uniqueresearch.substack.com/p/the-cooler-it-looks-the-more-likely
On-site reading page: https://ffcap.cn/en/research/the-cooler-it-looks-the-more-likely