MaHui Entrepreneurs | Booster Robotics' Cheng Hao: Not Chasing Trends, Just Building an Embodied AI Company That's "Still Here in Ten Years"

• This article is republished with permission from LatePost (ID: postlate); author: Xue Liang.

On June 6, as the 2026 World Cup kicked off, Chinese Foreign Ministry spokesperson Mao Ning reposted a video of robots playing soccer on X with the caption "Team China is ready."

The robots were Booster Robotics' next-generation humanoid T2. Though the company hasn't been around long, its CEO Cheng Hao has been working on robotic soccer for 20 years — since enrolling in Tsinghua University's Department of Automation in 2006.

Throughout our interview, Cheng didn't dwell on grand narratives about large models. Instead, he kept returning to Booster Robotics' conviction: layered deployment, data first. The company wants to get robots running in real-world scenarios first, using commercial deployments to accumulate data and create a flywheel effect. Cheng rejects aggressive bets. Humanoid robots in factories, he believes, show no path to scaled deployment; hardware margins will keep compressing. Algorithms alone can't build real moats, and embodied large models remain five to ten years away from maturity. Rather than waiting for castles in the air, better to build an operating system and practical applications now.

This is a company in no hurry to embrace grand narratives or chase capital trends. At a time when nearly everyone in the embodied intelligence industry is swept up by fundraising, valuations, and buzzwords, Cheng says financing is a two-way street. He'd rather move half a beat slower than the industry than spin stories or deviate from long-term strategy. What matters more: whether the business model works, whether products actually land, whether the organization stays grounded and resolute.

In May, Booster Robotics closed nearly RMB 1 billion in its Series A round. The latest tranche was led jointly by the Beijing High-End Manufacturing Industry Fund, Jingguosheng Fund, and Huakong Fund. The shareholder roster also includes Source Code Capital, Shenzhen Capital Group, Beijing Artificial Intelligence Industry Investment Fund, Beijing Robot Industry Development Investment Fund, Inno Angel Fund, Beyond Times, and IDG Capital.

From entering robotics through Tsinghua's automation department to building a cash-flow-healthy company two decades later, delivering thousands of physical units, and targeting education as a potential new multi-billion-dollar market — Cheng wants to prove that slow is fast. Rather than sprinting for temporary scale, better to build a company that still exists in ten years.

This, in Cheng's view, is the clearest lesson the mobile internet era offers the embodied intelligence cycle: find the business model first, then talk about changing the world.

1

Robotic Soccer: Twenty Years Sharpening One Sword, Validating the Minimum Closed Loop

LatePost: You were already working on robotic soccer 20 years ago. How did it all start?

Cheng Hao: I wanted to build robots since I was a kid. In high school, someone told me the automation department was where robots were made, so I applied to Tsinghua's automation program. At the time, Tsinghua had only two labs working on bipedal robots. The precision instrumentation lab shut down my sophomore year, leaving ours as the only one on campus.

Robotic soccer, to use current terminology, is the only truly deployable embodied agent. It requires locomotion (running, jumping, kicking), real-time decision-making (when to shoot), and multi-agent coordination (5v5 teamwork). The rules are clear, winning and losing unambiguous. Basketball depends on hands — too difficult to implement from 20 years ago to now. But soccer and walking both use feet, which relatively simplifies a complex problem.

LatePost: Your new T2 launched alongside the World Cup. Obviously its soccer skills are much better, right?

Cheng Hao: T2 is our flagship model for high-dynamic, high-explosive, complex-demand research scenarios. It's reached new heights in degrees of freedom, single-arm payload, computing power, and secondary development capabilities. Foreign Ministry spokesperson Mao Ning posted a T2 soccer video on X the other day — its autonomous kicking speed rivals world-class golden boot players. That was part of our World Cup content. During this World Cup period, we'll use T2 and K1 for a series of penalty kick challenges and interactive events.

LatePost: You kept competing in soccer tournaments after founding the company in 2023. What changed?

Cheng Hao: In 2024, I felt a bit disappointed. Nearly a decade had passed, yet everyone seemed to have made no fundamental progress. A German team once dominated the field, then open-sourced their algorithms. Everyone studied them for ten years and remained at the same level. Especially since everyone was still using servos — poor algorithm performance, weak capabilities, limited productization potential. It's like building cars: you can't use a three-wheeler engine, right? And servos were expensive.

But in 2025, our K1 and T1 competed like that German team back then, crushing virtually all opponents. We replaced servos with quasi-direct-drive joints, used cutting-edge algorithms, walked stably, weren't afraid of collisions, and shot precisely. Our biggest win was 20:0. T2's capabilities are even better. It's positioned as flagship-level, so computing power is maxed out, and the price range can cover broader users.

The sub-1-meter K1 can strike the ball with power. The ball hits the net, then the wall with a thud. Offline events have to restrict participant age to prevent injuries. This robot costs just RMB 39,900, with 22 degrees of freedom and 48 TOPS entry-level computing power. You can buy it on JD.com.

LatePost: After winning at robot soccer, what's next?

Cheng Hao: Beat humans. The goal is to beat humans by 2050; the mid-term target is beating U9 youth academy teams by 2030.

LatePost: But robot soccer is neither a commercial sport nor a productivity scenario. Doesn't that seem absurd?

Cheng Hao: Over 20 years, only one humanoid robot scenario has survived: soccer. Dancing, archery competitions — all gone. The moment humanoid robots play soccer well and beat humans, they'll have proven deployability. Soccer involves the full pipeline of locomotion, perception, and decision-making. Playing directly against humans also requires safety algorithm optimization for human-robot interaction. All of this can be closed-looped on the soccer field. It's the minimum viable product for validating cutting-edge algorithms. Very fundamental.

Like ByteDance: started with Neihan Duanzi, then Toutiao, then Douyin. Fundamentally, it's content production-distribution-consumption. Same algorithm, different carriers. The most vertical, simplest carrier can still perfect the algorithm.

LatePost: So soccer is embodied intelligence's MVP?

Cheng Hao: Yes. Testing in home scenarios: first, it's expensive; second, it's dangerous. No family would dare use it directly. But without validation, you can't get real-world data to iterate algorithms — you're locked out. Like autonomous driving: what if the whole world forbids it on roads? You need to get the car running first; data comes naturally, driving the embodied flywheel.

Plus soccer has entertainment value, especially effective for middle schoolers and below. At a recent Wukesong event, robots could already shoot through end-to-end neural networks. Elementary schoolers generally can't defend against penalty kicks — the force is strong, and they find angles. By mid-year, robots will run, dribble, and pass, further boosting entertainment value.

LatePost: Besides soccer, are there other scenarios that can achieve this kind of technical validation closed loop?

Cheng Hao: I can't think of any right now.

2

Layered Deployment: Why Build the OS Now, Not Bet on End-to-End

LatePost: Twenty years on, what's fundamentally different about embodied intelligence now versus before?

Cheng Hao: The broad logical framework is the same, but algorithms in each module keep upgrading. At the decision layer, for example, the current logic uses large models for generation.

LatePost: That still sounds like layered logic, different from what many companies advocate as end-to-end models.

Cheng Hao: The endgame is end-to-end solving everything, but that's still very far away. Even large model involvement at the decision layer has many problems. Soccer scenarios demand fast response — large models can't handle that yet; decision trees plus reinforcement learning are faster. For non-high-speed-response tasks, like guided tours, large model decisions can deploy. But that's still perception → decision → execution, not end-to-end solving everything.

LatePost: So Booster Robotics isn't working on embodied large models at this stage?

Cheng Hao: We don't bet on "castles in the air" end-to-end large models. But we're doing layered model deployment while collecting data from deployed scenarios — this is the Tesla path.

Embodied large models are extremely, extremely far from deployment. First, the paradigm isn't even known yet; from VLA to world models, it's still debated. Second, the training data doesn't exist. Even if you tokenize multimodal data, how much compute is needed? You've seen language models' compute demands. Not to mention we don't have this data at all, and we don't even know what data embodied models actually need.

All roads lead to Rome, but we're still at the stage of not having shoes.

LatePost: Which matters more, data or algorithms?

Cheng Hao: For human survival, is eating important or drinking? Both. But right now, we don't know how to do either.

LatePost: But many people talk about data approaches — real robot data, teleoperation, simulation data.

Cheng Hao: Teleoperation data produces demos where unscrewing a bottle cap looks like threading a needle. This is slow-thinking training fast-thinking — unreasonable. Real robot data and simulation data have similar problems. They can produce demos, work well in certain cases, but they're all transitional solutions.

Like in AI vision: OpenCV recognizing objects seemed impressive, but multimodal large models are the ultimate solution — nearly 20 years of iteration in between. A company can't wait 20 years, or even one or two. So we've kept deploying in scenarios like robot soccer, crawling through mud. Large models are flying in the sky; you need to crawl out first before you can fly up.

LatePost: How long before flying up?

Cheng Hao: From now, fastest five years, probably ten.

LatePost: Other companies on the embodied large model path — ten years to deploy?

Cheng Hao: Conversely, if no one pushes forward now, it might take longer. Like OpenAI with large models: initially not considering deployment, raising lots of money and betting resolutely on transformer, from 1.0 to 2.0 to 3.0, finally emerging with good results.

LatePost: Who's reliable enough to do that?

Cheng Hao: Probably major tech companies. Requires tens of billions of RMB investment. OpenAI isn't a major tech company, but bound to Microsoft, it has major tech company talent density and capital, so it could bet boldly. Embodied large models should be built now, but not by Booster Robotics.

LatePost: Then what is Booster Robotics building?

Cheng Hao: We're building a path complementary to embodied large models. They need data, need validation based on physical bodies, need deployment. And Booster Robotics is developing robot bodies, operating systems, and tools.

We're like Apple or Microsoft, building bodies, edge systems, and tools; large models are like the internet or cloud. Both types of companies will be needed in the future embodied industry. But for a startup to say it can do both and do both well — unrealistic.

3

The OS Era: Building Agent Ecosystems and Data Flywheels

LatePost: What's the prerequisite for developing an embodied operating system?

Cheng Hao: Software engineers first spend a year developing soccer-playing robots, becoming robot agent developers. Soccer-playing itself is an agent; through development, engineers understand what kind of systems engineering robots need. In this world, people who understand robots don't really understand software engineering; people who understand software engineering completely don't understand embodied intelligence. But after a year of playing soccer, we've cultivated engineers who understand both.

LatePost: What specifically does the operating system do?

Cheng Hao: Analogous to Windows and computers. General things — development tools, environment configuration — the OS integrates. GUI fundamentally lowers developer barriers. Previously, embodied development required a master's from a top school; with our tools, high schoolers who know Python can get started.

Eventually, an embodied agent ecosystem will develop. Not the familiar large model agents, but embodied agents — people combining different algorithm capabilities, plugging into the OS, solving real problems. Whoever builds the agent ecosystem becomes the biggest winner.

LatePost: You're about to launch a development tool called Booster Studio, which sounds like an "armory" for developers.

Cheng Hao: Yes, this should be the world's first tool software built specifically for embodied development. It has a complete simulation environment built in, from simulation to real robot one-click deployment.

LatePost: This reminds me of your "minimum closed loop" concept.

Cheng Hao: Right, we're about to host a global 3v3 robot soccer simulation competition. Developers can train their AI agents in Booster Studio, submit to the cloud for automatic battle and scoring. Winning strategies can be directly deployed to real K1 or T2 robots. This opens our 20 years of real soccer field validated closed loop to developers worldwide — no need to own hardware, you can experience the full process "from code to goal."

LatePost: Can I understand this as Booster Robotics trying to build the Android system for robots?

Cheng Hao: Yes, decoupling applications from hardware. Our OS can adapt to various chips, models, brands. Upper-layer developers only focus on developing agents, running on Unitree robots or Booster Robotics robots alike.

LatePost: Won't hardware become hard to sell this way?

Cheng Hao: Embodied intelligence is entering the late stage of the body era. Hardware technology has converged, becoming increasingly standardized, margins very low. How much difference is there between buying a Lenovo and an HP laptop? Not much.

LatePost: Operating systems seem like something only major companies can build?

Cheng Hao: Did Apple and Microsoft succeed first then build operating systems, or succeed because they built operating systems? Actually the latter. The ecosystem from an OS is a very strong moat.

LatePost: What qualifies Booster Robotics to build an operating system well?

Cheng Hao: We're the team most suited for this. From day one, the team split into two groups: classmates doing automation and bodies, and internet development engineers — nearly half with purely software backgrounds.

Globally, Booster Robotics is the robotics company that most values software engineering. Mention embodied intelligence and everyone thinks hardware, algorithms, without realizing that before truly achieving embodied large models, the long OS era requires extremely complex software engineering. It's dirty work, massive code volume, but builds real moats. Historically, no company has ever built moats through algorithms alone. Large models have proven this once again.

LatePost: Won't AI lower software engineering barriers?

Cheng Hao: Don't you think that actually helps a startup like us grow into a giant? (laughs) At a major company, a project used 300 people; now maybe 30 can do it.

LatePost: Internet-background engineers don't understand embodied intelligence. How can they be architects?

Cheng Hao: First let them go play soccer. People who understand both are extremely rare in this world. But after a year of soccer, we've cultivated a batch.

4

The Education Market: A Multi-Billion-Dollar "Apple II Moment"

LatePost: You seem quite to-C. Small robots available on JD.com.

Cheng Hao: One of our directions is to go small first, landing faster and commercializing. Currently, the highest shipment volume is still small humanoid robots. In the small humanoid robot field, we're number one in shipments. More peers will enter this year.

LatePost: Who's buying these small robots?

Cheng Hao: Two products: T1 at 1.2 meters, focused on research needs — locomotion control, navigation, multi-agent decision research. K1 is smaller, for education and teaching scenarios, cheaper, defined as an entry-level embodied development platform.

Education is a huge market. In the 1980s in the United States, 1990s in China, massive computer lab construction played an important role in popularizing computers, creating China's engineer dividend. Early on, Apple actually sold large volumes to schools.

Embodied intelligence will go through this phase. Robot body plus operating system plus development tools can replicate the commercialization of the Apple II or DOS-era computers. Many students want to learn robotics but can't afford it; they rely on school computer labs. Booster Robotics fits this scenario very well: software-wise, the OS has already lowered developer barriers; hardware-wise, small robots aren't cumbersome, very safe.

Research or soccer tournaments have small total addressable markets, but computer labs in education — that's a multi-billion-dollar market.

LatePost: Landing in China involves complex relationship management?

Cheng Hao: Our first robot sale was to an overseas customer. In 2025, 40% of Booster Robotics' revenue came from overseas; in the first two months this year, it rose to 60%.

This is a truly solid business model. Compared to the somersaults everyone else is chasing — no business model there.

LatePost: What about robots in factories?

Cheng Hao: For a business model to work, robots must be cheaper than humans. That's not possible now. Simple things in factories, robotic arms can do. Complex things that robotic arms can't handle, robots also can't do, or cost much more. Like car assembly — robots can't handle flexible wiring, still need humans in the end. The math doesn't work out; better to just hire people. I see zero opportunity for the business model of robots landing in factories.

LatePost: Home scenarios?

Cheng Hao: Having done this for so many years, watched for so many years — not optimistic. Home scenarios are too complex. Toys or chairs on daily routes — how to navigate around? A bed against the wall for folding clothes — how to reach? Putting in cabinets — how to open doors? These complex situations greatly reduce success rates. The instability of bipedal robots, easy to fall — this will actually be well solved this year.

LatePost: How does your locomotion control compare to Unitree?

Cheng Hao: Actually about the same. Algorithms can't build moats. Unitree's real moat is high-explosive joints, very deep hardware accumulation.

LatePost: Having sold 1,000 units, any supply chain learnings?

Cheng Hao: Started using third-party parts rather than self-developed. Gradually shifted to self-developed as volume grew. Supply chain in China isn't hard; besides hiring the right people, it's about controlling volume — when capacity is one unit per week, sell one unit externally. A client wanted 100 units; I directly said no.

Because our mass production gradually ramps up: from one unit per week, to three per week, to ten per week, to maybe fifty per week — step by step, not suddenly claiming we'll produce 100 per week or 1,000 per week.

Our commercial strategy advances in parallel. Early on maybe sell one unit; now can sell five. This is exactly the process of production and commercialization advancing alternately, step by step. The benefit is that in early stages, you can use relatively low costs, hire fewer people, and first stabilize one-unit-per-week capacity.

5

Endgame Vision: OS-Era Data Assets, for Future Large Model Latecomer Advantage

LatePost: If major companies now build embodied operating systems or even large models, how do you respond?

Cheng Hao: Major companies will find it hard to commit to embodied models now because there's no clear implementation path. The person in charge might face poor performance reviews for years to come; no one wants to do it. Even if they do, it might shorten model maturity from 15 years to 12 years.

For major companies to succeed, the number-one leader must personally do it. But currently, no number-one is willing to personally oversee embodied large models, because the technical path hasn't converged. We call this TPMF — technology product market fit.

LatePost: But once they do, it's disruptive for you.

Cheng Hao: Exactly. In the mobile internet era, software was suppressed. Windows used to sell expensively, now it's free. But Microsoft, with accumulated capital and talent from before, caught up again in the cloud era and large model era.

LatePost: One battle after another, at different dimensions.

Cheng Hao: In the body era, we were already preparing for the OS era. In the OS era, of course we prepare for the embodied large model era. There's possibility for overtaking on the curve. ByteDance only started large models in 2023, but Doubao is very successful now.

LatePost: The prerequisite is first becoming ByteDance.

Cheng Hao: Exactly. I often tell investors: when we have very stable billions in revenue, hundreds of millions in profit, we'll definitely maintain a very large team to do embodied models. Again: algorithms have no moats; you can achieve latecomer advantage.

Through the long OS era, via operating systems and development tools, we can obtain the most data — meaning the opportunity to train the best models. We don't do large models in the air, but we crawl through mud, crawl out, then fly up.

LatePost: What's the biggest lesson from the mobile internet cycle for the current embodied intelligence cycle?

Cheng Hao: Find the business model first, then talk about changing the world.

6

Business Choices: Rather Than Valuation, Building a Company That Still Exists in Ten Years

LatePost: You seem quite conservative in how you describe commercialization.

Cheng Hao: Yes. I think finding investors is itself a two-way selection process. Investors are like dating — you don't deliberately try to persuade the other side, make them like you. My thinking is: communicate quickly, find the one who truly clicks. Finding investment is the same — the core is to very resolutely explain our logic, then quickly find investors who recognize this logic.

That said, our commercial growth momentum is very strong. In Q1 this year, cumulative shipments grew 500% year-over-year; one quarter's volume matched the first eight months of 2025. January-February new signed orders rose 800% year-over-year.

We've already established solid commercial self-sustaining capability. We can sell wherever we go. Whether at CES or the Yizhuang Marathon, robots we brought were snapped up on site. This is the most direct proof — we can really sell, really generate blood, completely without needing to spin stories or hype concepts to attract investment.

LatePost: Aren't you anxious? Looking at this market environment, everyone seems a bit crazy.

Cheng Hao: I might feel anxious inside, but we'll still very resolutely use logic to explain the reasoning. Second, our fundraising progress is actually okay — basically completing three to four rounds every year. It's just that overall we're not so aggressive. We won't tell big stories, quickly pull investors in, sign high-risk post-investment terms, then rapidly inflate valuation.

Our overall pace is about half a year behind the industry's valuation increase speed. This actually has no impact. But we definitely don't want to cater to investor preferences by spinning stories, doing what investors like — that would have very, very big impact on the company's strategic direction.

LatePost: How do you view the intensity of 2026 market competition?

Cheng Hao: I've always held this view: as long as your product and business model are fine, it's impossible not to get investment. This market will ultimately vote with its feet, investing in enterprises that achieve commercial success.

LatePost: But you have to survive until then.

Cheng Hao: Right, but the core is that our business model itself works, and overall shipment volume continues growing several-fold, several ten-fold. This isn't simply surviving. For example, if you rapidly inflate valuation now, go public quickly, but without good business model support, you'll fall eventually, and then the whole team faces huge upheaval.

Honestly, at that point everyone stops focusing on business and stares at the stock ticker daily — mentality gets heavily affected. Our logic: if persisting in this direction for ten years can build a company like Microsoft, why rush to some level in these two years? I don't think that's the most fundamental thing.

LatePost: There's widespread anxiety that not getting money means dying, or subsequent funding won't be as easy.

Cheng Hao: Whether money is easy to get does follow objective patterns. But whether a company dies depends on whether you have a product, whether that product has a mature business model.

LatePost: Among investor questions, what's the hardest one for you to answer?

Cheng Hao: Haven't encountered one yet.

LatePost: Is hiring difficult?

Cheng Hao: Somewhat. Though on the other hand, talent supply itself exists — it's just many candidates care about whether company valuation is high enough. This is actually a good thing. In my experience, people like this, once hired, likely harm the organization — they often focus on these surface things daily, or care about company publicity momentum, not fundamentals, not whether they can do their own job well. Many directions make quick money more easily, easier to do superficial sales or analysis work. But our field still needs very grounded people who recognize the direction and can resolutely accumulate bit by bit.

So we never spent money on PR before, because it felt meaningless. Until this Spring Festival, when industry competition environment deteriorated. But I think this is actually a good thing, equivalent to forcing us.

(Forcing us) to fill in some capabilities, make them robust. The core is I feel that relying solely on product technology isn't enough now, because consumer-level to-C opportunities haven't formed yet. Money thrown out now can only land on brand, and brand subsequently can't sustain it — the public is very forgetful, this information has little value. Need to be more pragmatic, invest funds in R&D, in building business models.