Is AI Hardware the Next Consensus? From Robot Fighters to an All-Purpose Cyber Badge

“The point is not to run AI inside the hardware. Instead, AI should define or control the hardware.” Wang Le, on his understanding of AI hardware

Guests: Wang Le | Founder of FoloToy; Zhen Di | Founder of Hunban Technology

In conversation: Raymond | Technology investor; Mark Tang | Founder of Caidazi AI

AI Passport looks like an employee badge. At the time of recording, its community had already created more than 1,300 ways to use it, while the company itself had developed only around 10. Some people install games, some use it as an exhibition pass, and others make gifts for their children. Wang Le picks out community games to play himself. When he likes one, he also tracks down its creator to ask how it was made.

With hardware, you usually finish the features before putting it on sale. Wang Le runs the numbers in reverse: the share defined by the manufacturer falls from 95% to 85%, and might eventually reach just 45%, leaving the rest to users and AI. At first, it sounds like selling an unfinished product. Yet the appearance and quality still have to meet consumer-product standards. What users get to change are the functions inside. The hardware has to be properly made; the company can make fewer decisions about its functions on the user’s behalf.

There were robots and robot dogs at the WTCC recording venue. Our other guest, Zhen Di, has moved from AI toys into robotics applications. His problems are more specific: when a robot dog arrives at a dock, it needs to plan its own route and carry the right modules for inspection. Customers are buying patrol and inspection capabilities, and they weigh the tasks against the costs.

Why would someone pay for a small device that still needs configuring after purchase? Once a robot can walk, talk and box, what makes a customer buy another one? Wang Le says users flash different functions onto the same hardware; Zhen Di says his team actively looks for use cases that support repeat purchases. Both have to look at what happens after delivery: do users keep creating, and do the machines actually get work done? When discussing AI hardware, I think those questions deserve more attention than another item on the spec sheet.

The full conversation follows, edited and abridged.

I. Can Hardware Go on Sale with Only 65% of Its Functions Defined?

Raymond: There are already so many AI apps on our phones. Why buy another piece of hardware?

Wang Le: I have kept thinking about what form AI hardware should take. I now describe one approach as “user-defined AI hardware”: users employ AI to define their own hardware in a more personalized way.

When we reviewed our products from the company’s founding to today, we realized we had been doing this all along. At first, we did not have a complete, ready-made product. We made a circuit board, prepared the software, and let users take it home and DIY. We also made toy-conversion kits: undo a few screws, replace the motherboard in an existing toy, and it gains new functions.

People asked whether this was like Arduino. We said it was even more basic than Arduino. We prepared the accompanying software; users could take it home, plug it in and start their own DIY project. Some storytelling toys they already had at home could also be modified by replacing the motherboard. We started with a single circuit board.

Later, with self-hosting, we supplied the hardware and software while users handled configuration and setup. Looking back, the business has not fundamentally changed. AI’s capabilities have. Modifying a product or writing a program used to require substantial technical skills. Now, if you explain what you want, the right AI agent can do a lot of the work for you.

Mark Tang: There are two kinds of products on the market. One has its functions already defined: users buy it and use it. The other, like AI Passport, depends on users deciding what it should do. I originally thought the second kind would be hard to get started with, yet many people are willing to buy your product. How did you arrive at this approach?

Wang Le: I am a consumer too. I have bought glasses and all kinds of AI products. They do not necessarily meet all my requirements, and modifying or hacking them used to be very difficult.

Standardized products can meet most of a particular group’s needs, but there have always been people who want personalization. As AI lowers the barrier to modifying products, that group will grow. Our business has not changed since the day we first had the idea, but our user base grows as AI becomes more capable.

That is why I want the predefined portion of our hardware to shrink gradually. We might start by defining 95%, then 85%, and eventually go below 50%, delivering a product that is 45% defined and leaving the remaining software and server configuration to users.

I use “users” broadly. They might be individual developers, small businesses or large enterprises. They can take our hardware, add suitable AI and infrastructure, and build a complete solution for their own setting. Standard products will still exist. What we want is for people to use AI to write software, then use that software to customize the hardware’s functions.

Raymond: Let me clarify this. If you can ship at 65%, how do you actually measure 65% versus 45%?

Wang Le: In terms of the product definition we do ourselves, we would like to be able to sell at 65%. After receiving it, users still need to configure it a little and set it up the way they want.

We once tried selling this approach to enterprise customers, who found it hard to accept. We then tried making consumer products ourselves and found that it could work. Take a toy: do we have to integrate every music and content service? We can leave interfaces available and let users connect what they need.

We have to bring the hardware’s appearance and quality up to consumer-product standards. The functions inside can then be customized to users’ needs and their children’s needs. Every household is different, so the resulting product will be different too.

We used to put AI inside toys to make a conversational product. Now we are handing part of product definition over to users as well. I have become more confident about this. A few of us finish a product meeting, and AI has already started development. The points we discussed can quickly turn into a demo.

Zhen Di: I have known Le for quite a while. When we first did a proof of concept for AI toys, our online research turned up his public development logs. We later discussed technical details such as speech recognition, speech synthesis, communications and latency. But very early on, he was emphasizing AI-native products and trying to give users as much of AI’s own capability as possible.

People now talk about embodied intelligence. I prefer to think of our work as “giving intelligence a body.” Hardware used to be mainly about augmenting human capabilities: glasses, headphones and phones were designed around people. In the future, some hardware may also be intended for our AI assistants, giving them a body and the ability to sense the physical world.

An assistant needs to record things, hear sounds and understand its surroundings to build context. When people and AI use these devices together, the definition of the product naturally changes.

II. Can Someone Who Cannot Code Provide Technical Support on Site?

Mark Tang: AI software and agents keep improving, but one step is still missing: devices that collect context. Meeting recorders, glasses and devices worn on the wrist or around the neck help AI understand what we do and what we want to do every day. With more context, a personal assistant can help more proactively.

Zhen Di: Software is changing too. People used to open different apps and press different buttons. Now we are gradually moving from GUI to LUI. Some products are starting to let AI skip those operations and solve the problem directly.

One distinctive thing about AI Passport is that it is not fully fixed in its final form. In the past, when developing a product, people’s needs would not change drastically over a short period. AI now iterates so fast that by the time you finish everything, you may already be behind. Letting more people participate in defining the product is one response.

Wang Le: That is why, when people say this is “the last Programmers’ Day,” I think we can put it differently. Companies may stop treating programming as a separate occupation, but everyone can be a programmer, developer or creator.

At an event in Beijing a few days ago, the person we sent to provide technical support was no longer a technical specialist. It was our product lead. He does not know how to develop software at all, yet he could solve technical problems on site. That would not have been possible before.

Mark Tang: Hackathon participants are getting younger too. First they were programmers with some work experience, then university students, and now teenagers are making things themselves. As AI capabilities become widely available, will the barrier to mastering these tools keep falling?

Wang Le: Yes. But we have also seen something counterintuitive: older people account for a substantial share of those buying and using our products. Some are more than ten years older than I am.

They are learning programming tools and vibe coding to use at work or to make things for their children. Some parents have actually used these tools to build a hardware product for their child, then excitedly shared it in the group. They feel that they made it themselves, and that gives them a real sense of achievement.

Young people may accept AI more naturally as part of life. Older people can feel a more urgent desire to change. At least among our buyers, the shift is clear. The age range of our users has not been narrowing.

Raymond: Apart from you, who else is making hardware that users define themselves?

Wang Le: I think many manufacturers will try once the model has been validated. But being able to do it takes time and accumulated experience. We already have users, fans and a community of some size who are willing to try. If another hardware manufacturer releases a product with unfinished functions or a low level of completeness, users may not accept it.

We have a group of people willing to build with us. That is the basis for trying this model. The product itself also has to be designed to make participation easy.

Raymond: So how do you get users willing to try making something?

Wang Le: Show them what others have already made. Let them try it, then copy and modify it. If copying a prompt lets them make something, it becomes easier to keep going.

We want to offer a product that is easy to get started with, gives people a sense of participation, and ends with something they can hold and play with. Both the process of creating and the result matter. Once a user succeeds, that sense of achievement encourages another attempt.

Other hardware is moving in this direction too. Some products allow functions to be extended through MCP or an SDK, which also lets users help define them. The question is how much the manufacturer does and how much the user does. For some products, having 90% predefined is enough. Not everyone needs to be as open as we are.

Raymond: I think hardware is developing a long tail too. People used to have their own websites and apps. Now even a very small need can have its own piece of hardware: install software you wrote yourself, and it runs.

Wang Le: Some users already buy several units and flash different functions onto the same hardware, carrying whichever one they like. It does not have to integrate everything into one screen the way a phone does.

III. Why Did the Company Make Only Around 10 of More Than 1,300 Uses?

Raymond: Your community of ways to use the device sounds a little like an app store: some people make things, and others use them. Among registered users, how many actually make something themselves?

Wang Le: We have estimated that creators make up about 5% of registered users. It is quite a small proportion. Beyond registration numbers, we can see downloads, comments, uploaded creations and other community data.

We do not call it a UGC community. Our community is 100% AIGC. Users participate by creating through AI. Without AI, the creator group would be much smaller and might not sustain this model at all.

I checked before coming here today. There should now be more than 1,300 publicly available uses, while we have developed only around 10 ourselves, all very simple demos. Users created essentially all the rest.

Raymond: Out of those 1,300-plus uses, has anything made you think, “You can even do that with it?”

Wang Le: So many things. I install games on the hardware myself and choose ones that are well implemented and work well. Some rhythm games are remarkably complete.

When I find a creator like that, I try to meet them, ask how they designed it, and encourage them to keep going. They do not have to build only on our platform. They could try making something for iOS or Steam too.

I have seen many things users made that I feel I could never have thought of or built in my lifetime. Even everyone at our company put together might not have managed it. Everyone has different educational, life and learning experiences. You only realize someone could come up with something like that when you see it and try it.

This platform and these tools give me a chance to hold other people’s ideas in my hands and play with them. It feels like connecting to someone else’s context.

Mark Tang: I see many users sharing spontaneously on social platforms. Beyond the ways they use it, people also seem eager to show the physical things they have made.

Wang Le: Once you have finished a piece of software, sharing it often means dealing with app stores, servers and other hurdles. Once I finish a little hardware device, I can simply give it to a friend or hand it to my child at home.

It turns AI’s output into a tangible object. That feels different from generating a small app on a phone. Users feel that this thing was made specifically for them.

Phones are certainly versatile. A phone from several years ago could already meet many digital needs. But phones now also carry advertising and all kinds of things built into their systems. This small device can have no ads and no push notifications. A user can choose a single function to keep and concentrate on it.

Some users say hardware that looks limited in its functions actually helps them enter a flow state. In the AI era, we often have several agents and conversations running at once, moving from one thing to another, and sustained focus is difficult. Some people like realizing an idea gradually on this little device.

Raymond: You play with the creations and pick them out yourself. That already sounds a lot like running a content platform.

Wang Le: Yes. I select many of the official recommendations too. I contact creators, discuss how they implemented things, and try to turn their experience into something other people in the community can use.

I both define the product and use it. I keep trying things to see how we can lower the barriers further and how good uses get created.

Raymond: It sounds a bit like Lego? You buy something partly finished, and the user decides what it becomes.

Wang Le: A little, but I cannot find a perfect one-to-one comparison. It is somewhat like Nintendo, somewhat like Lego, and like a combination of other companies too.

AI brings substantial changes. I think it is hard to define a company of the next era with a single sentence from the previous one.

IV. I Barely Notice When I Record Others. Why Am I Careful When They Record Me?

Raymond: AI needs to collect context to understand me. But if I wear glasses or put out a recording device, does the person sitting opposite want to be recorded? How do you handle that?

Mark Tang: I first look at whether a device can blend into everyday life without requiring attention. If you already wear glasses, another device you have to decide whether to take each day adds a mental burden. Putting the function into glasses you already use is more natural.

Some glasses remove the camera, reducing both weight and intrusiveness. Wearing glasses with a camera used to make me feel a little strange, and it could make the other person uncomfortable too. Some devices are now difficult to recognize from their appearance, yet they may still be continuously collecting conversations.

A device may be convenient to use, but it still changes how you feel when talking to someone.

Zhen Di: We considered this a lot when making children’s toys. Some products synchronize all of a child’s conversations to the parents. The safety rationale is that parents can trace and check them. We deliberately removed that feature.

We also debated cameras, eye displays, computing power and battery life. In the end we removed some things and kept functions such as storytelling and flapping wings. For conversation records, parents cannot see the full exchange. They can see the week’s keywords.

Those words can be a bridge between parent and child, helping parents understand what interests the child or whether something is troubling them. We blur out the specific things said.

Personally, seeing glasses or a recording device opposite me also affects what I talk about.

Raymond: I do not notice it at all when I record other people, but I am very careful when someone records me. Many people have this contradiction. You see yourself as recording your life. When someone else records you, you start thinking about what you should and should not say.

Wang Le: Devices with cameras bother me more. If I know someone is filming me, I may not want to say as much. Recording prompts an instinctive reaction too: when meeting a friend I have not spoken to in a long time, I first check whether there is a recorder around their neck.

We joke that you record yours, I record mine, and we are even. But the concern remains.

Zhen Di and I agree on some of this. We did not put cameras in our children’s products. For audio recording, some products require users to trigger it actively and make it clearly perceptible that recording is happening, rather than letting it happen quietly.

In a work setting, when meeting clients, you can let them know first that you want to write a summary afterward. That is usually easier to accept. I mind more in private conversations, outside public settings, especially with someone I do not know very well. The result may be a narrower conversation or less depth. There is a trade-off.

V. AI Wants to Save Users Time. Will Platforms Agree?

Mark Tang: AI phones try to cover many things: ordering food, booking rides, shopping and handling documents. Once it becomes a phone, there is a temptation to want it to solve every use case.

But China’s internet ecosystem is not that open. Many actions get blocked. A product demo may say a service is integrated, while the capabilities it can actually invoke differ from what users understand. Each platform has its own boundaries, making it hard to connect ecosystems and data properly.

In the future, some specialized providers may be willing to work with operating-system vendors to improve their competitive position. Once one partnership opens up, others may change too. But in the short term, setting a very high ceiling for a product’s scope and capabilities does not necessarily make it work.

Wang Le: We have also kept thinking about what AI phones and personal-agent hardware should look like.

A large part of both the mobile internet and the PC internet consists of advertising systems. They get people to look at things and spend time, while also providing efficiency. AI, however, helps users save time and may even remove the need to look at ads. Existing platforms want to occupy users’ time. That creates a major conflict.

If someone’s business model and monetization depend on that, they will not simply hand over their position because you have built a smarter entry point.

When making hardware, an obvious approach is to build a phone, get the operating system and underlying layers right, and assume the other problems will follow. Yet the form closest to a personal agent may not need a screen at all.

Put a phone motherboard into a small device, use a computer at home, and carry a phone or laptop when going out: you can already do a lot. In terms of efficiency and convenience alone, many hardware forms can solve the problem. The hard part is balancing ad distribution, time spent by users and efficiency.

So I do not think some current approaches to AI phones are already good solutions. More exploration is needed. Perhaps products and hardware for narrower uses will emerge first, meeting a specific set of needs.

Mark Tang: In some activities, the process itself is what the user wants. Some people enjoy browsing Taobao, just as they enjoy window-shopping. We might go to a mall already knowing which shoes we want, but many people simply enjoy looking around.

If AI replaces shopping and browsing, it may be a negative optimization. It saves a few steps but also removes something the user wanted to experience.

Raymond: I think we need to separate those things. If you already know what to buy, faster is obviously better. But if someone wanted to browse in the first place, how is removing the browsing an improvement?

Mark Tang: Exactly. Some platforms create content themselves because they want users to spend more time there. You cannot understand it simply as finding a product, placing an order and finishing. When AI does things for users, it enters an existing business system.

VI. Why Does Building AI Companions Turn You into a Content Company?

Raymond: Making Labubu talk does not seem technically difficult now. But should it talk?

Zhen Di: My answer is no. We have discussed this privately with friends, and relevant companies are already thinking about product definitions and standards for AI toys.

Some products emphasize cuteness, others how they feel to touch. Being able to build a function does not mean you have to add it. Companions and tools are different products.

Mark Tang: You do not want your pet at home to actually talk to you either. Its physical presence and feel can be comforting. When you are stressed, holding it feels good. I feel very comfortable holding my dog, but if it spoke to me, that would be quite frightening.

Zhen Di: Japan explored some of this quite early. People have talked about “weak robots.” Some products today still show traces of those early products, and some simply do not speak.

When evaluating products early on, we also asked whether they could make you want to care for them. Wanting to look after something is a different relationship from having it ready to complete a task for you at any time.

There is another practical issue. The further you go toward companionship, the more you discover that making the product work gradually turns you into a content company, an entertainment company, an IP company or a design company.

If you try working with major IP owners very early, you spend a great deal of time discussing things with their legal teams. No one has made this kind of product before, so contracts need many revisions. Both sides may agree on the direction, but there are still many practical things to resolve before moving forward.

Strictly speaking, the products we previously made with IP partners still followed the previous generation’s definition: functions and packaging were finished, users could open the box and use them, and what they could do was bounded in advance. A storytelling device can tell stories and hold conversations, but stories still account for a large share. Whether users keep using it also has a lot to do with the IP and content. The AI hardware contains those experiences.

Wang Le: We have kept discussing cooperation with some IP owners, but have not found a particularly good way to make it easy for both sides to proceed.

That is why I have never especially emphasized our toys as companion products. We emphasize function. A large plush toy is an AI storytelling device for children at bedtime. A desktop product can manage time or help parents give spelling dictation. A device taken to a museum can act as a guide. The purpose has to be clear.

Because they are fluffy, people group them with companion products. But to me, they are more like educational products or tools. I compare them with learning devices costing several thousand yuan or desktop time-management tools costing several hundred. I think about which specific function the user is paying for.

Users may later define and modify the product themselves and feel, “I made something. That feels great. I am happy.” That creates emotional value too. But for us, it is an attribute that comes with making a tool.

We have always wanted to use AI to create a personalized tool with an acceptable cost of use.

VII. If Big Companies Fill the Jar with Soybeans, Which Gaps Can Startups Fill?

Raymond: Model companies and big internet companies are all moving into hardware. How do you avoid going head-to-head with them?

Wang Le: I do not pay much attention to how any particular big company makes hardware. I generally do not study their direction or strategy.

Raymond: Do you look down on them?

Wang Le: No. I cannot do what they do either. We let users define hardware. Big companies may not want to do that, because the initial volumes can be very small.

People’s needs vary enormously. Perhaps only ten, twenty or a hundred people need a particular piece of hardware and its accompanying software. A big company might study it, decide the market is small and will not make much money, and pass.

I think big companies put soybeans into a jar. We make the little mung beans that fill the gaps. Solving very specific needs well is enough. We do not have the money or energy for platforms and foundation models. We take care of small things, one by one.

Mark Tang: Use your comparative advantage.

Wang Le: Exactly. I also try good products. But what those companies do does not determine which direction we must take.

Raymond: Zhen Di also has robot-body manufacturers to contend with. How do you compete with them?

Zhen Di: I do not think the starting lines are that far apart right now. Things are changing extremely fast at this stage. Talent concentration and funding differ, of course, but those conditions can have side effects too. It is not easy for a big company to validate and implement something new either.

We have enterprise business, and often we work with big companies while doing the actual implementation ourselves. The relationship can be complementary.

I also do not think the future must produce a single huge hit as in the previous period. As capabilities converge, we may see more different products. Some underlying capabilities and infrastructure will still produce very large companies.

For individual products, something that grows organically can be more resilient. Changes to core personnel, organizational structures and project staffing in large teams all affect whether a project survives. A big company’s project does not necessarily have a higher survival rate than a small company’s product.

Raymond: But in China, someone can take your hardware apart and study it. If they make something similar and cut the price, what do you do?

Wang Le: We worked through those possibilities before launching. First, you cannot simply assume someone else can definitely sell at half our price. If they rely on heavy subsidies to push the price down, they also have to consider whether that is sustainable. Our supply chain and pricing competitiveness are not weak in the domestic market.

But if you look only at hardware specifications and performance, we have no advantage at all. Other hardware manufacturers can take it apart and understand it.

Are Nintendo’s consoles the strongest if you look only at hardware? No. We want to build a user community, an ecosystem and a gathering place for ways to use the device: a place where people create content with AI.

Building a company is not about getting everything done at once or becoming an overnight hit. You have to run it carefully, step by step. If we can see growth every month and every year, and it meets our expectations, we keep going.

Look over a longer period. After a few years, if the company is still at the table, still growing, still releasing products the market needs, and has its own user community, that is quite an achievement.

Mark Tang: Brand, community and users’ perceptions are the most important assets.

Raymond: After individual users buy the hardware, do they have to pay service fees?

Wang Le: Those follow-on services for individual users are free. We are also exploring ways for developers to make money, so people who create good uses can earn a return.

We also have enterprise business. Exhibitions, company events and hackathons use our hardware. It can act as a pass and work with software for check-ins, games, points and interaction, letting the same card serve different functions in a particular setting.

What individual users create and what enterprise customers ask for both show us what else the product can do. You cannot see those things by comparing hardware spec sheets alone.

VIII. Once Robots Can Perform, Who Pays for Them to Work?

Raymond: It is quite a leap from AI toys to robots. How did you get here?

Zhen Di: The starting point and way we collaborate have basically not changed. Our earliest idea was Cyber Partner, Hunban: giving AI a body so it could interact with the physical world. A robot may be one eventual form, but there is a process to getting there.

We chose toys early on because the technology then could deliver some clear functions and create experiences that were relatively hard to replace. Even though the finished product reached consumers, we built it with an IP partner. There was still enterprise-to-enterprise cooperation behind it.

Our robotics applications now are also based on enterprise partnerships. Robots can easily cost hundreds of thousands of yuan, and robot dogs may cost tens of thousands. Asking ordinary consumers to buy one just to explore is still a high barrier. Among early customers for embodied-intelligence applications today, enterprises are more common.

Businesses look at ROI and at what problem the robot solves. We work with robot-body manufacturers and accumulate experience and capabilities through specific projects. The question has not changed: how do we use AI well, bring it into real life and real settings, and create value?

Raymond: Companionship, security, inspection: which do you do first? How do you choose?

Zhen Di: We are trying two directions. Companion products have relatively high tolerance for errors and address needs that are somewhat easier to reach. But competitors rush in quickly. Consumers find it hard to tell products apart, and some products still have problems with the actual experience, retention and returns.

The other direction is a narrower use case that requires step-by-step validation but solves a clear problem. A robot might embody a brand or act as a guide, interacting and telling corny jokes. People can accept that it cannot do many other things.

As robot dogs enter work settings, tolerance for errors gradually falls. We mainly look at three types of tasks, called the 3Ds in English: dull, dirty and dangerous. Where people do not want to do something, or are not suited to doing it, a robot’s value becomes more concrete.

Our team started working on companionship fairly early and accumulated some experience. But proving the product is better than someone else’s has always been a problem. The enterprise business also grew gradually from there.

Raymond: Who makes the dog itself, the sensors on top, and the control system?

Zhen Di: Lingshao is our payload platform and central control platform. The robot bodies come from different manufacturers. When people see a dog, they tend to attribute every capability to that body. But walking, sensing the environment and completing a particular task need to be considered separately.

Payloads are the sensors or modules added for a task. We work with robot-body manufacturers and different module suppliers to connect those pieces. Whether the software works and whether the devices cooperate are all part of delivering the application.

Robots have locomotion intelligence, interaction intelligence and task-execution intelligence. Moving and talking with people do not mean a robot can already do a particular job well. Task-execution intelligence is what is most lacking now, and it is a long-term undertaking.

IX. Can One Customs Project Be Replicated at the Next Port?

Raymond: Robots are expensive. A customer buys one. Then what? Which use case will make them buy another?

Zhen Di: We work backward, actively looking for use cases suited to repeat purchases, testing them and selecting the genuinely good ones. In inspection, there are already cases and customers operating at scale and making repeat purchases.

We are also building the Lingshao platform to lower deployment barriers. Users should not have to understand an entire embodied-technology stack, including algorithms, motion control and interaction, and then connect all the hardware before they can start. For the same capabilities to enter more real-world settings, deployment must be easier.

Mark Tang: Could you talk specifically about the customs application, or the one at West Bund?

Zhen Di: We have worked on the customs project for quite a while, going through testing and POC stages. We use robot dogs to replace or assist some patrol work at ports and docks.

Cargo stacks at docks change quickly, and the environment is not ideal. Devices need to plan routes and navigate on their own. Containers have spent a long time at sea. When opened, there can be all kinds of odors inside. Some involve hazardous materials. Inspection can be risky, and it is not always practical for a person to go into every location first.

To help with this, a robot dog first needs basic capabilities such as autonomous navigation. It is more like a mobile platform. Its form need not always be strictly limited to a dog. What matters is that the algorithms work in the physical environment to plan and move.

The central platform can monitor devices and control them as a fleet, connecting them to one another. Above that, different payloads are added for different tasks.

I compare these payloads to AI’s eyes, ears, nose and mouth. The eyes might be vision systems or lidar, enabling it to see and record. The ears can make judgments from voiceprints, noise and vibration signals. The nose requires working with specialist module manufacturers for gas detection and similar tasks. The mouth can make announcements during security and patrol work.

You cannot assume those capabilities exist simply because the body is a robot dog. You have to look at which modules are installed, who the partners are, and which solution addresses which problem.

Raymond: What used to be a set of recognition and monitoring capabilities is now installed on something that moves, so it can go out and do a specific job.

Zhen Di: Yes, it becomes more concrete and practical. This customs case is national in scope and can be replicated. Patrol settings offer similar opportunities. We look for projects where completing one creates opportunities to reuse the work elsewhere.

Overseas markets also calculate the economics around tasks. With RaaS, you have to consider which task is solved, local labor costs, the social structure and restrictions on use. We can pilot in China first, but scaling still requires adaptation to different environments.

Wang Le: We tried selling overseas. After evaluating compliance and related issues, we felt that going abroad amounts to starting a company a second time. It is not just selling the product there and being done. Domestic and overseas work have to proceed together, and products will differ for different markets.

Competition in China is intense, and development is fast. We want to keep doing the work here well while completing the preparations needed overseas. The two have to move forward together.

Zhen Di: Cyber Studio, which we are also building, is another attempt to lower the barrier to entering this process. Even without buying a robot for now, people can DIY their own robot or cyber partner on the platform and start training it in simulation.

We also entered EngineAI’s URKL robot boxing competition and had just reached the round of 16 at the time of recording. We have kept accumulating the relevant technology and team capabilities. The products may appear to serve different groups: some help enterprises deploy robots, while others let developers create. But much of the underlying technology is related.

We need to let more people participate, use AI and experience what it can do. Without that mindset, this would be hard to keep doing.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.