Unitree CEO on China's Robot Revolution
“Humanoid robots will revolutionize every industry within our lifetime"
Hangzhou-based Unitree Robotics is among the top players in China’s robotics industry, developing best-selling civilian quadruped and humanoid robots.
Unitree’s H1 humanoids captivated over a billion people with a traditional folk dance performance during China’s 2025 Spring Festival Gala. Just weeks later, Unitree’s CEO was the youngest front-row participant in Beijing’s highly-anticipated private sector summit.
To better understand the Chinese robotics industry and where it’s headed, we’ve translated and annotated qan interview with Unitree’s founder and CEO, Wang Xingxing 王兴兴. Originally conducted in April 2024 by Titanium Media (TMT), the interview covers:
Why LLMs aren’t enough for the robotics industry, and why Wang predicts the emergence of a large-scale AI model for general-purpose robotics by the end of 2025,
Factors driving the global humanoid robot boom, and why China is uniquely poised to succeed in this industry,
The techno-optimist vision for the economy of the future, powered by humanoid robots as well as machines of alternative forms,
The timeline for mass adoption of AI-powered general-purpose robots,
Unitree’s strategy for competing against foreign and domestic robotics firms.
We’ve added some editorial notes for your enjoyment, including commentary by anonymous robotics PhD and current industry player KL Divergence.
Interested in learning more? For past ChinaTalk coverage, see Angela’s work on China’s leap into industrial robotics and China’s humanoid robot industry.
The Translation
Original Article | Archive | Title: “Dialogue with Wang Xingxing: Humanoid Robots Will Reshape All Industries Within My Lifetime” | Author: Rao Xiangyu 饶翔宇 for Titanium Media (TMT) 钛媒体APP | Editor: Zhong Yi 钟毅
On February 17, 2025, a highly anticipated private enterprise symposium was held in Beijing.
At the event, the presence of Wang Xingxing, the founder of Unitree Robotics and a post-90s entrepreneur, attracted market attention. Wang was seated in the front row among the business representatives, alongside industry giants such as Zeng Yuqun, Jack Ma, Ren Zhengfei, Wang Chuanfu, Lei Jun, and Pony Ma. Among them, Ren Zhengfei, Wang Chuanfu, Liu Yonghao, Yu Renrong, Wang Xingxing, and Lei Jun delivered keynote speeches.
As a startup company, Unitree Robotics and Wang Xingxing have experienced what can be described as a “rocket-like leap” in growth.
[KL Divergence: Perhaps. Though consider that Unitree has been an established player, particularly in quadruped robots, for 10 years. They have worked hard and scraped their way up inch by inch by serving the small markets which existed for their product (mostly researchers). Their recent prominence has come from riding the wave of interest in humanoid robots by creating low-cost, easy-to-use, but not particularly advanced or capable, humanoid robots for research.]
Public records show that Wang Xingxing earned his bachelor's degree from Zhejiang Sci-Tech University. Due to poor English proficiency, he failed to gain admission to Zhejiang University for his master's studies and was instead placed at Shanghai University. In an interview, Wang once mentioned, “During my three years in high school, I only passed English exams three times in total.”
From 2013 to 2015, while pursuing his graduate studies, Wang, despite having limited resources and funding, independently designed hardware and control algorithms and combined them with industrial motors to develop the robotic dog XDog. This project won second prize in the Shanghai Robotics Design Competition. After graduating, Wang embarked on an entrepreneurial journey focused on robotic dogs.
[KL Divergence: Actually, much of the hardware and control algorithms were from publicly-available robot and actuator designs published by Western researchers, such as Sangbae Kim (MIT) and Dan Koditschek (UPenn). What Unitree really excelled at was (1) iterating high-performance actuators, robot designs, and research-grade control algorithms, and (2) leveraging the Chinese supply chain to create low-cost, high-performance, highly-reliable combinations of these key technologies. In other words, they productized the research.]
Unitree Robotics, founded in 2016, initially specialized in the development of quadruped robotic dogs and successfully sold them worldwide, becoming one of the leading players in the industry in terms of product shipments. By 2023, the company ventured into humanoid robotics and quickly became one of the most closely watched companies in the field. In 2025, Unitree Robotics’ latest humanoid robot appeared on the stage of the CCTV Spring Festival Gala, garnering widespread public attention.
In April 2024, Wang Xingxing, the founder of Unitree Robotics, had an exclusive interview with Titanium Media APP. The relevant content can be found below.
Wang told TMT APP that the fundamental reason behind the humanoid robotics boom is the emergence of large AI models. Previously, it would take one to two years for a humanoid robot to learn to walk, but now, with AI algorithms, this can be achieved in just one month.
Regarding the future development of humanoid robots, Wang expressed strong optimism. He believes that by the end of 2025, at least one company worldwide will have developed a general-purpose robotic AI model. This foundational model, he explained, is like a complete set of building blocks, with large language models being just one piece. Other crucial components include visual perception, tactile sensing, decision-making, and interaction systems.
Looking at an even longer timeline, Wang told TMT, “Within our lifetime, humanoid robots will be able to revolutionize every industry, from manufacturing and agriculture to services and industrial sectors. Taking it a step further, governments could potentially deploy 100,000 humanoid robots to construct an entire city. On a smaller scale, robots could even shrink down to the size of cells, transforming all aspects of our natural environment.”
Below is the full interview between TMT and Unitree Robotics founder Wang Xingxing, with slight editorial adjustments.
“The Turning Point for Humanoid Robots Has Not Yet Arrived”
TMT: A few days before our meeting, Boston Dynamics, a star company in the robotics field, announced that its hydraulic-powered humanoid robot would be phased out, and future developments would focus on electric-powered products. What are your thoughts on this?
Wang Xingxing: Boston Dynamics has been making robots for many years, and they’ve also been working on commercialization for a long time.
As for hydraulic drive systems, I had already believed before 2013 that this approach could not be commercialized. The reason is simple: it relies entirely on precision mechanical components, and once you involve such components, costs will never come down. Moreover, all hydraulic systems leak oil. That’s why you hardly see hydraulic systems in consumer vehicles anymore — they’ve all been replaced by electric drive systems.
So, if Boston Dynamics wants to continue developing humanoid robots, switching to electric drive is definitely the right path. The only surprising thing to me is that I assumed around 2018 that they had already started working on an electric version. But later, when they had made no detectable progress, I just stopped paying attention.
TMT: Compared to hydraulic systems, is electric drive better suited for large AI models?
Wang Xingxing: Compared to hydraulic systems, electric drive is all advantages and no disadvantages. As for whether electric systems are better suited for AI, that’s harder to say. However, electric drives have lower production costs, offer greater motion flexibility, are safer, and also lighter in weight.
TMT: Now that Boston Dynamics has switched to electric-powered robots, combined with their existing training data, do you think they could iterate faster than competitors in this new wave of competition?
Wang Xingxing: It’s hard for me to say. However, we remain quite confident, because we’ve been working on quadruped robots for many years, and a lot of the algorithms and components we’ve developed can be directly applied to humanoid robots.
Another important point to note is that most of the top AI talent in the U.S. isn't at Boston Dynamics—they’re at Google, NVIDIA, and OpenAI. Boston Dynamics' strength likely lies in hardware development and traditional humanoid robot control systems.
TMT: So, would you say that the emergence of large AI models is a major turning point for humanoid robots?
Wang Xingxing: I don’t think we’ve reached that turning point yet.
Right now, it’s more like a starting direction. There's a common misconception—many people think that large language models like ChatGPT can be directly applied to robots, but in reality, that’s not the case.
TMT: Why not?
Wang Xingxing: Because LLMs aren’t designed for robotics in the first place. ChatGPT operates purely on text logic, and its entire training dataset is based on text data. It doesn’t perform well in robotic environmental perception — this is a global challenge, not just a problem for one company.
While the humanoid robotics industry does use AI, the technology is actually very different from large model technology.
TMT: But some companies have claimed that large AI models can already recognize different types of plates, allowing robots to identify and pick them up.
Wang Xingxing: That’s not something we can verify. It was just a video, and no one has confirmed its authenticity.
Besides, there’s no data proving that if you swap the plate for an apple, a pear, or something else, the robot would still be able to recognize and handle it correctly. Personally, I don’t see any evidence of real technical breakthroughs coming from Silicon Valley — it still seems quite conventional (中规中矩).
TMT: So, large AI models are not the key turning point for humanoid robot development? Are they less important than people think?
Wang Xingxing: The models themselves are not important for robots, but the underlying technological direction they represent is very important.
Right now, large models are mainly focused on language models — but no one has yet developed a true large-scale model specifically for robotics.
TMT: That brings us to the big question — what triggered the humanoid robotics startup boom in 2023?
Wang Xingxing: The reason is really quite simple: Tesla started working on humanoid robots.
Elon Musk has already disrupted industries like automobiles and rockets, growing them into massive sectors. Now that he's entering humanoid robotics, governments and various institutions want to get started early, rather than waiting for Musk to succeed first and then trying to catch up.
[KL Divergence: I think this is a little bit just-so and playing to the audience a bit too much. The fundamentals are more important. Elon and Optimus is definitely the spark which ignited the wildfire. But the kindling was years and years of slow and steady progress on batteries and electric motor power density made it finally possible to create practical (as in, strong and light enough) electric humanoid robots, around 2020-21. Elon's team caught on to this a little early, because these are also technologies that Tesla happens to to be deeply invested in. But others were doing it already, just quietly.]
At the same time, ChatGPT and other LLMs have expanded the public’s imagination of AI’s potential. You could say these models ignited excitement and enthusiasm across the industry. Right now, what we’re seeing is just the beginning — the momentum will only grow stronger.
As hardware and AI technology advance each year, the impact of humanoid robotics on the world will be massive and transformative.
“It’s simple, really not as complicated as most people think”
TMT: Current large AI models are just the beginning. What's the future direction of the industry or where should everyone's efforts be focused?
Wang Xingxing: There are many directions. The first step is adapting AI for robots - developing robotic vision, perception, understanding, execution planning, and various operations.
I'm excited just like everyone else. I personally feel this industry will develop rapidly, including robots, large models, and AI. I believe by the end of 2025, at least one company globally will develop a relatively general-purpose robot large model.
Our company hopes to develop it ourselves, but realistically speaking, the probability is higher that an American company will achieve it first.
[Angela: Wang is optimistic. Depending on what goalposts you set, training a robot “foundation model” requires large, multimodal datasets that take time and capital to collect or synthesize – including vision, sound, touch, motion, social and environmental interaction, and so on. In some ways, Wang’s prediction aligns well with the Chinese government’s stated goal of mass production of humanoids by 2025 and world leadership by 2027. At the same time, he realistically recognizes the strength of US innovation. The main takeaway here is that to Wang — and likely to many of his industry counterparts — the humanoid robot race is accelerating towards some decisive moments.]
TMT: So that brings up the question of open source versus closed source.
Wang Xingxing: If we develop it, it definitely won't be open source.
TMT: Is there a unified model between robot large models and robot dogs?
Wang Xingxing: Most robot dogs are implemented through reinforcement learning, which is a relatively mature technology.
Robotic large models or robot world models can be applied to all robots, not necessarily humanoid or dog-shaped ones - they're universal tools. I've always believed that robots don't necessarily have to be humanoid; the humanoid shape is just one of many possible forms. I've never insisted they must be humanoid.
TMT: But the mainstream view is that humanoid forms are better because our whole society is built for human-shaped frames.
Wang Xingxing: They might like to say that, but I've never believed it.
You can build entirely new physical worlds. Why would you need a humanoid form for mining? Why would you necessarily need a human shape for building houses? Of course, humanoid forms are important, or relatively important, but they're not everything.
For example, at home, people might prefer humanoid robots for performing scenarios or accompanying you on trips. But for building houses or transporting things - physical labor - there's no need for them to be humanoid. Plus, humanoid forms might give people a sense of owning a slave if you make them do unpleasant work, making their owners feel uncomfortable.
TMT: Would you feel sorry for them?
Wang Xingxing: Current AI hasn't reached that level yet; it can’t perceive such things.
But if its AI could perceive pain or negative emotions, then yes, that might be problematic. But there’s no need to feel sorry now, because it’s still just an inanimate object [死物, literally a ‘dead thing’] with limited intelligence.
TMT: I'm curious about something — even though their intelligence is limited, when you push them, why do they display human-like staggering movements?
Wang Xingxing: Because that’s what the AI was trained to do through reinforcement learning.
TMT: So it’s imitating human behavior?
Wang Xingxing: Some behaviors aren't imitation; they’re determined by natural laws. You could say physical laws constrain these robots to move in certain ways. If an alien had a human shape, its movement would probably be the same as well.
TMT: Currently people break down robots into cerebrum, cerebellum, and the physical body. What's your view on this?
Wang Xingxing: I’ve never liked separating the cerebrum and cerebellum so distinctly. One model is enough - why divide it into two? I don't think it's necessary.
Of course, there might be various modules within the model, but overall I prefer treating it as a single model. From walking to fine-grained operations, we implement everything using AI in a completely end-to-end manner. From visual perception to leg execution, one model handles it all - no intermediate mathematical formulas whatsoever.
TMT: Can the hardware capabilities keep up?
Wang Xingxing: For robots, it's just a few joints — it’s really not that hard. Just sensors feeding into the model, and then the model outputs to the joints. That's all.
TMT: Your understanding of humanoid robots seems simpler than others'.
Wang Xingxing: It is simple, not that complicated.
TMT: For example, others might think dexterous hands are difficult for fine operations because they require more accurate recognition and finer motion control.
Wang Xingxing: It's very difficult if you use traditional technology, so you can't rely on traditional approaches. Without technological innovation, there's no point in working in this field. Of course, you can't express it so directly — better not to go too far beyond public understanding, otherwise I'd probably get cursed out (骂死).
TMT: What specifically do you mean by non-traditional?
Wang Xingxing: It's new AI, end-to-end. It means not having to manually write lots of software programming rules in between, nor perform traditional image recognition.
TMT: How do you implement that?
Wang Xingxing: Modify the model. The underlying AI is the same, but your entire model structure and algorithms are different. I can't explain this too specifically - it would be hard to understand. For example, you don't need traditional image annotation or image understanding at all. You can input images and videos into a model, and the output is directly the robot's joint trajectories, then you just train it. You can still do image annotation, like labeling images of apples. But annotation has only one function: interacting with humans, helping it better understand people. For the robots themselves, there's no difference between an apple and a pear.
TMT: Compared with the mainstream opinions, your logic and industry judgments are unique.
Wang Xingxing: The mainstream viewpoint still has many issues. As a startup, if your thinking is just mainstream, it just won’t work out well for you. You must see the development direction for the next few years, and once you see it, plan ahead accordingly - then you're certain to win, or at least not lose. If you only see what everyone else is talking about, others can certainly do better than you - how could you stand out?
TMT: In your view, what will the next few years look like?
Wang Xingxing: I can't get too specific, but what's certain is that the industry will progress extremely rapidly.
TMT: How fast are we talking here?
Wang Xingxing: It's basically beyond imagination. The current pace of AI deployment in factories — globally, technological progress is extremely fast and has almost proven viable.
TMT: Currently, no company can fully utilize robots for work.
Wang Xingxing: But the entire logic has almost been proven. This doesn't mean robots can do everything, but work-capable, end-to-end robots are nearing maturity. A more general-purpose robot model will likely be developed by a company globally before the end of 2025.
TMT: That fast?
Wang Xingxing: It could be even faster. Some people have already seen where this is going - though it sounds a bit boastful, I feel I've seen it too. Following this direction, with some additional time, manpower, and money, it can basically be developed.
“All technological breakthroughs have a large element of luck"
TMT: What specifically does a robot model refer to?
Wang Xingxing: You can think of it this way: first, it has strong mobility capabilities applicable to most terrains, possibly with some mobility skills exceeding humans. For instance, its obstacle-crossing ability, speed, jumping ability might be better than humans. Another aspect is working in factories - it can do many tasks without requiring manual programming. Through large model capabilities, with just a little teaching, it can learn by itself and then perform well.
TMT: Is simulation training in virtual environments still necessary?
Wang Xingxing: Probably not all that necessary. Once you've trained it well and validated it, you don't really need simulation anymore. Of course, completing the hardware won’t happen right away, but I think that's just a matter of time. As for AI, there's still some uncertainty. Although I just said I'm personally optimistic it will emerge before the end of 2025, it might not happen - it could take 3-5 years before it's developed. It depends on humanity's collective luck - sometimes it just comes down to luck.
TMT: How do you understand this kind of luck?
Wang Xingxing: Many technological breakthroughs depend on luck. For example, if Einstein hadn't existed, someone else would probably have discovered his theories. But it might have been delayed by several years, or even decades. You can consider that all technological breakthroughs have a large element of luck involved.
TMT: Another point: besides algorithms and models, large models need data. Is data collection currently very difficult?
Wang Xingxing: There are indeed many things that need to be done, but there are methods for addressing them. It's not as complicated as people think - many problems aren't as complex as people imagine. You know, in all current technology fields, if you really look, there's nothing truly complex; everything is relatively direct and simple. Even
TMT: So is your industry also divided into two camps - optimistic and pessimistic? For example, you're more optimistic, thinking the whole thing isn't that difficult.
Wang Xingxing: It definitely requires time and intellectual investment, but these are things that can be solved and advanced. They’re not like room-temperature superconductors or controlled nuclear fusion. The biggest problem with room-temperature superconductors and controlled nuclear fusion is that there’s a question mark over whether they’re physically possible. The universe might simply not allow such things to exist, and humans might never achieve them no matter how much time and effort we invest. Artificial intelligence robots are common things, not something extraordinary — just the intelligence of a bunch of humans and animals. Intelligence is a widespread phenomenon. Some animals are very smart and can understand much of what humans say, they just can’t speak. And crows — some crows can even use tools directly. So, intelligence doesn’t have many limitations or physical constraints; it can be replicated.
TMT: What’s the biggest motivator for your work?
Wang Xingxing: To be honest, what moves me personally is AI.
A few years ago, an investor asked me whether our company would ever develop humanoid robots, and I told him, “We would never do it, even if it kills us.”
[KL Divergence: Great honesty here. It's true. Virtually the entire field considered humanoid robots a hopeless tarpit, which would consume all of your time and money and render not progress. Even in robotics research, humanoids were a quirky backwater reserved for the cranks and over-optimistic.]
Back then, humanoid robots were far too complex. Traditional algorithms simply couldn’t handle such intricate machines. The conventional approach to training humanoid robots relied on highly skilled engineers manually writing mathematical equations to model movement. These equations would then be solved to determine the robot’s motion trajectory. But this method had severe limitations—if the environment changed, the equations often became invalid, requiring entirely new models to be designed from scratch.
This approach also led to an overwhelming amount of code, and as the system grew more complex, it became nearly impossible to maintain manually. However, AI changes everything. As long as the model is well-structured and continuously fed with data and compute, AI can iteratively optimize itself through trial and error. By leveraging reinforcement learning and reward mechanisms, AI can automatically retain successful training outcomes and discard ineffective ones, dramatically improving training efficiency.
Recent progress in AI technology—both in capability and speed—has far exceeded my personal expectations. That’s why, despite having worked on humanoid robots for just over a year, our performance is already exceptionally good. The reason we’ve been able to move so quickly is simple: thanks to advancements in AI.
The benefit of AI is that once you’ve built a strong model, the rest is just a matter of compute—you don’t have to manually fine-tune everything. If you need to test a scenario, OK, all you need to do is feed the system more data. This is also why Tesla’s autonomous driving team is significantly smaller than Chinese autonomous driving teams. I know for a fact that Tesla’s team has only a few hundred people, whereas some companies in China have teams numbering in the thousands.
TMT: Is this also why newer players have been able to surpass Boston Dynamics?
Wang Xingxing: Exactly. If we were competing with Boston Dynamics purely using traditional algorithms, we wouldn’t stand a chance. The reason is simple: Boston Dynamics has an entire team of PhDs from MIT, and there’s no way China could outmatch them in that domain.
TMT: Looking ahead, what do you think will be the key differentiator among humanoid robots?
Wang Xingxing: Robotics is an integrated product. Unlike fuel-powered vs. electric vehicles, where the underlying technologies are fundamentally different, the differentiators in humanoid robots will be more incremental—primarily in specific engineering optimizations, such as motor scale, motor placement, workspace dimensions, structural design, and leg configurations.
The same principle applies to AI. Take large language models — they’re fundamentally pretty similar. The biggest points of differentiation are in the details rather than in fundamental design; OpenAI’s GPT architecture is still relatively clean.
“In our lifetime, humanoid robots can reinvent all industries and the natural environment.”
TMT: Commercialization is also important. How can startups survive in an increasingly competitive landscape?
Wang Xingxing: The business logic is very simple. As long as your product is better than your competitors’ in all dimensions, then you will profit. What remains is the question, how big is the entire market? Right now, our company has a strong market position, so we have captured most of the easily accessible revenue opportunities.
TMT: What do you mean by ‘easily earned revenue’?
Wang Xingxing: From having high shipment volumes. We sold quite a few quadruped and humanoid robots last year.
TMT: How many did you sell?
Wang Xingxing: It's hard to say exactly, but it's under a few hundred. However, we definitely sold the most in the domestic market.
TMT: Who bought them?
Wang Xingxing: A variety of buyers, including research institutions, AI companies, and businesses pursuing real-world applications.
TMT: How can you move so fast and manage to sell your products?
Wang Xingxing: Because we have a strong foundation. There’s significant overlap between robotic dogs and humanoid robots. Our company holds advantages in technical R&D, AI algorithms, manufacturing, and sales channels. We already have an established customer base and ready-to-market products. Other companies have to build everything from scratch, which takes time.
TMT: Is your revenue sufficient to support R&D?
Wang Xingxing: Our company maintains healthy gross profit margins, complemented by ongoing funding.
TMT: For humanoid robot startups, is the ability to secure funding a core advantage?
Wang Xingxing: It’s hard to judge the industry right now because it’s too hot. Many companies with basic foundations have raised some funds, which are at least enough to keep them afloat.
There’s certainly no shortage of funds in this industry. When we started, we were poor. Compared to back then, things are completely different. Now, some companies have been around for only a year and already have a valuation of 1 billion yuan. It's astonishing. The industry isn’t short on capital, and neither are they.
But I think that before the industry truly takes off, having too much money is pointless. It’s difficult to allocate effectively, and if spent indiscriminately, it could easily be wasted. At this stage, neither the technology nor the business models have been fully validated, so throwing money around wouldn’t be wise.
Take bike-sharing, for example. It worked because the business model made sense. Once that’s proven, the only thing left is scaling up, and there’s nothing left to do but pour in funding.
TMT: What do you mean when you say the technology and business model haven’t been fully validated?
Wang Xingxing: It means that neither the technical framework nor the commercialization strategy is fully developed. Even if you have the capital, you don’t necessarily know how to deploy it effectively.
TMT: What are the main technical challenges?
Wang Xingxing: For humanoid robots, the biggest question is how to integrate with AI models—we don’t have a definitive answer yet.
TMT: Another observation—most humanoid robotics entrepreneurs today are quite young. (Wang Xingxing is from the ‘90s generation.) Why is that?
Wang Xingxing: It’s simple. Older generations just aren’t as interested in this space. AI technology is evolving at an unprecedented pace and older knowledge is becoming outdated – knowledge of the technology we had five years ago is practically irrelevant. The younger generation is fastest at learning and applying the new advancements. Traditional internet startups had a low barrier to entry — basically anyone could become a product manager. But humanoid robotics isn’t a conventional industry.
[Angela: We’ve written before about how this generation of emerging technology creates space for young, enthusiastic talent to make an impact — DeepSeek is a good example of this. It would be interesting to know if, like DeepSeek, Unitree draws its success from China’s pool of homegrown talent. Wang Xingxing himself never studied or worked abroad.]
TMT: Earlier, you mentioned the potential for a breakthrough innovation. Were you referring to how humanoid robots and AI models can be integrated?
Wang Xingxing: Yes, more or less.
TMT: But aren’t AI models just modular components that can be put together like building blocks?
Wang Xingxing: The differences in AI models go far deeper than that. Take Transformer architectures, for example—there are still endless ways to optimize and refine them. Researchers are even exploring alternatives to Transformer-based models altogether. The AI field is full of opportunities for technical breakthroughs, and there’s still vast room for innovation.
I anticipate that by 2025, we’ll see a significantly improved AI model for general-purpose humanoid robots. When that happens, industry momentum will accelerate even further, to the point where companies from around the world try to enter.
TMT: At that point, do you think hardware or software will be the first to breakthrough?
Wang Xingxing: Software will be the key driver. No matter how advanced the hardware is, without the right software, it’s just an expensive pile of metal.
TMT: So given the current pace of development, as soon as the right software emerges, the hardware will be able to keep up?
Wang Xingxing: Absolutely. Hardware won’t be a bottleneck. If it’s really needed now, we can scale production quickly by aggressively deploying capital. If necessary, we could push manufacturing capacity to its limits — pay engineers 10 times their normal salary, work around the clock, and purchase all the necessary equipment. With sufficient investment, we could have mass production up and running in as little as a few months to a year.
TMT: How does China’s hardware capabilities compare to those of other countries?
Wang Xingxing: China has a significant edge in hardware. The cost-performance ratio is much higher.
TMT: Why is that?
Wang Xingxing: First, in the U.S., hardware development doesn’t receive as much attention—most of the top talent is focused on software. Second, manufacturing and labor costs in the U.S. are much higher than in China.
TMT: It seems like they are prioritizing software, while our strength lies in hardware.
Wang Xingxing: Exactly. Most major U.S. companies focus primarily on software. But at Unitree Robotics, we are developing both software and hardware, because maintaining competitiveness requires full-stack capabilities. As a relatively smaller company, we can’t afford to focus on just one domain. Large corporations can get away with specializing in software and outsourcing hardware, but for us, abandoning hardware development would be an unwise strategic move.
TMT: Why has robotic dog technology matured faster than humanoid robotics?
Wang Xingxing: One reason is that robotic dogs have been in development for a longer period, and their form is more stable. They don’t require complex dexterous manipulation, like grasping and handling objects.
Another key reason is that, in the past, there was a much larger community of developers working on robotic dogs, whereas today that number has declined. In AI, the maturity of a technology is often directly correlated with how many researchers are actively working on it.
For example, large language models have advanced much faster than AI for robotics simply because more people are involved in developing them. Ten years ago, computer vision—especially facial recognition—was in its golden age because so many researchers were working on it. Image-based AI took off because it was relatively easy to experiment with; all you needed was a decent computer.
But robotics is a different story. It requires hardware simulation and real-world testing, which makes it much harder for individuals to participate. That’s why the field has been slower to progress. However, as I mentioned earlier, the industry is now accelerating because a growing number of people are entering the space. More minds working on a problem naturally lead to faster breakthroughs.
TMT: Does Unitree Robotics have a clear product roadmap and timeline?
Wang Xingxing: We will launch new products every year.
TMT: What do you envision for Unitree’s next-generation robots?
Wang Xingxing: The next generation will undoubtedly surpass current models in every aspect—appearance, performance, AI capabilities, and more.
TMT: Can you give a specific example?
Wang Xingxing: Our goal is for humanoid robots to perform real industrial tasks—working in factories, assisting in production assembly, and handling logistics.
TMT: Do you have a release timeline for the next generation of robots?
Wang Xingxing: It’s not convenient to disclose at the moment.
TMT: Unitree Robotics has already completed eight rounds of funding. Do you expect fundraising to accelerate moving forward?
Wang Xingxing: I think we’ll be fine. As the industry gains more attention, we’re seeing increased interest from investors.
TMT: What do you think will be the ultimate future for humanoid robotics?
Wang Xingxing: In the future, humanoid robots could redefine entire industries—from manufacturing and services to agriculture, mining, and construction.
I imagine a distant future in which governments could deploy tens of thousands of humanoid robots to build entire cities from the ground up. At that point, infrastructure is fully automated, housing is provided at no cost, and people no longer need to work because robots sustain the entire economy. That’s entirely within the realm of possibility.
Also, right now when we talk about humanoid robots, we picture machines that are roughly human-sized. But in reality, humanoid robots could build smaller robots, and those smaller robots could build even smaller ones. This process could continue indefinitely, leading to robots at microscopic scales.
Eventually, we might see robots as small as biological cells. Who knows what’ll happen then? What we perceive as bacteria could actually be tiny robotic entities. The entire natural environment could be restructured from the ground up. When that happens, governments will need regulations to prevent unchecked proliferation, or these robots could consume resources uncontrollably.
[Angela: A very science-fiction vision indeed. But Wang’s fantasy of a robot economy resonates with Beijing’s investment in industrial robotics as a path for economic advancement. ChinaTalk will continue tracking such developments in robotics.]
TMT: Do you think we will see this level of technological advancement in our lifetime?
Wang Xingxing: Absolutely. The only missing piece is AI. Once AI breakthroughs happen, everything else will follow naturally.
This will fundamentally reshape the world. I’ve always believed that when we look back at today’s society after the emergence of general-purpose AI and humanoid robots, it will feel as distant and primitive as looking back at the Stone Age.