"Current Robot‘s Embodied AI Remain Inadequate": Unitree CEO's Speech at the 2025 World Robot Conference, Exclusive Interview with People's Daily and NetEase Technology

Recently, Unitree Robotics CEO Wang Xingxing attended the 2025 World Artificial Intelligence Conference in Beijing and delivered a speech titled “Opportunities and Challenges in Scaling the Robotics Industry.”

Wang noted significant growth in the robotics industry in the first half of this year, with overseas giants continuously driving development. Current robot hardware is sufficient but embodied AI remains inadequate, resembling the stage before ChatGPT’s emergence, with a critical breakthrough likely in 2-5 years. The industry overly focuses on data while neglecting model issues, with VLA models showing limitations. Video generation models for robot control hold greater potential, while RL scaling laws need breakthroughs. Key future directions include end-to-end embodied AI models, low-cost durable hardware, and distributed computing. Industry progress requires global co-creation.

Wang Xingxing’s speech:

Hello everyone, it’s a great honor to share with you here. I’m Wang Xingxing, the founder of Unitree Robotics. Our company was established in 2016, so it’s been nine years, which is quite a long time. Initially, we focused on high-performance autonomous robots, and in recent years, we’ve shifted to high-performance general-purpose robots. Currently, our product lineup is quite diverse, with a wide range of models.

Let me give a brief introduction. From 2013 to 2015-2016, during my student days, I developed a robot called XDog. This robot was quite groundbreaking, essentially pioneering the technical solution for low-cost, high-performance autonomous robots globally. Before I developed this solution, most of the industry, including many universities, used industrial motors and servo drives, which were very expensive and bulky, making them impractical to use.

When I came up with this idea in 2013, I even considered dropping out of school to start a company, as the concept of dropping out to become an entrepreneur was quite trendy. But after some thought, I realized it wasn’t realistic at the time. It was just an idea—I hadn’t built anything, and I lacked resources. So, from 2013 to 2015-2016, I focused on developing this robot. You might not believe it, but the entire R&D investment for this robot was only about 10,000 to 20,000 RMB.

In 2015, I took this robot to a competition in Shanghai and won second place, earning some prize money. That was my first bucket of gold. When we founded the company in 2016, that prize money was our initial funding. This robot was quite representative, and many of the technical solutions used in today’s humanoid robots are very similar to it.

Currently, our company focuses on three main autonomous robot models. The first is a smaller robot, like the Go2, which we released in 2023 with an integrated large language model, making it highly functional. It comes equipped with a laser radar, 3D laser radar, language models, and various other features. Over the past few years, including this year, it has been the top-selling quadruped robot globally in terms of shipment volume.

Another model is the B2, which we released a couple of years ago, primarily for industrial applications. We want robots to actually work in various scenarios, including industrial ones. The B2 has exceptional continuous load capacity and battery life, which remain very strong to this day.

As for humanoid robots, I’ve mentioned on many occasions that for years, I was firmly against developing them. Back in 2009, when I was a freshman in college, my first robot was a small bipedal humanoid, which cost about 200 RMB. At that time, I realized that building humanoid robots was extremely challenging, both commercially and technically. So, for many years after founding the company, we avoided them. However, from 2021 to 2022, global AI development surged, especially with the advent of ChatGPT, marking significant AI advancements. Companies like those led by Elon Musk and other cutting-edge firms brought immense attention to this field, driving global interest in humanoid robots.

To put it bluntly, in 2022, we hadn’t even started working on humanoid robots, but clients were already placing orders and paying deposits without us having a product. This directly led us to start developing humanoid robots in early 2023. Our first humanoid robot, the Unitree H1, took just over six months to develop. Being our first generation, its appearance might be a bit rough, but its dynamic performance remains very strong. For example, it was this robot that performed at the Spring Festival Gala, and next week, it will compete in a robot competition in Beijing—everyone can keep an eye out for that.

Additionally, last year, we released our second humanoid robot, the G1. This robot is very representative. Since its release, many other companies’ new humanoid robots have started to look very similar to ours in terms of structure. The G1 has excellent mobility and flexibility. Priced at 99,000 RMB for the base model, it’s highly competitive. It’s likely the top-selling humanoid robot globally last year and this year.

If you browse Douyin or other video platforms, you’ll see many of our clients spontaneously livestreaming with this robot—almost daily. It’s quite visible and competitive.

Just a couple of weeks ago, we released our new humanoid robot, the R1. This robot is highly cost-effective, priced at 39,900 RMB for the base model, and it’s smaller, about 1.2 meters tall. Its standout feature is the customization potential—we’ve left plenty of room for clients to modify or paint its appearance. People love personalizing their robots with designs, clothes, or coatings. We expect this robot to be frequently seen globally this year and next. We have it on display at our booth, so feel free to check it out. Mass production isn’t finalized yet, so deliveries might take some time.

We also recently released a new quadruped robot, the A2. Its key feature is its light weight—only 35 kg, easy for an adult male to carry—and its strong load capacity, up to 25-30 kg continuously, with over six hours of battery life unloaded (about 20 km). Its design and perception capabilities are also very strong, and it’s been well-received since its release.

We’ve also updated our wheeled robots, which you might notice are more flexible. This is because we applied complex motion algorithms developed for our humanoid robots to our quadruped robots, significantly boosting their flexibility. These robots are industrial-grade, dustproof, and waterproof, making them suitable for industrial scenarios. Over the past few years, they’ve been used in 24-hour operations in industries like power grids and factories, with automatic charging, patrolling, and recognition of images, gases, or other elements.

Our larger industrial-grade robot weighs about 70-80 kg with wheels, which is a bit heavy, but its flexibility remains excellent. It can handle continuous loads of 40-50 kg—capable of carrying a person, though it’s not recommended due to safety concerns.

Many think we focus on robot performances or fighting competitions, but our core goal is to make robots work. We’ve done extensive data collection and open-sourced some of it—check our public repositories, where we release algorithms and data every couple of months for anyone to explore. Since 2016, we’ve developed our core components in-house, including motors, reducers, and some sensors.

For example, our second-generation 3D laser radar has a wide field of view (over 90° x 360°) and is very affordable, with a retail price of about 1,000 RMB. It’s standard on our base-model quadruped robots and offers 2-3 cm accuracy, suitable for both indoor and outdoor use. For low-speed robots like cleaning or logistics robots, this radar can reduce costs to just a few thousand RMB, making it highly competitive.

Over the past six months, our robots’ AI technology has advanced rapidly. For instance, a video from January shows our robot running somewhat stiffly, but recent OTA updates have made its movements much smoother, even on complex terrain like slopes or rocky paths. This feature is now available to all clients. Many customers, both in China and the U.S., use our robots for various events.

A notable example is our robot’s dance performance at this year’s Spring Festival Gala, which became a cultural symbol blending technology and tradition. This was a collaboration with director Zhang Yimou, who proposed the handkerchief-twirling and tossing movements, which we technically implemented. The biggest challenge wasn’t a single robot dancing—we achieved that last year—but having 16 robots automatically change formations and dance in sync. With 3D laser radars, they handled positioning and choreography autonomously, with no way to interrupt once the music started. The result exceeded everyone’s expectations, becoming a cultural icon, with cosplays and student enthusiasm following.

Our latest dancing and fighting technologies allow robots to learn any motion, but the sequences are fixed, requiring pre-collected motion data for AI training. This makes it less flexible, as you can’t adjust or interrupt the sequence, marking it as an older-generation technology.

Surprisingly, most of our AI team focuses on making robots work—handling tasks with dexterous hands or full-body operations. We’re not aiming for single-function AI, like folding clothes or cooking, but for general-purpose AI that can handle tasks like serving tea, factory work, or performances. This is a significant challenge, and our progress isn’t perfect, which is why we publicize this less.

Our latest technology, used in fighting competitions, is more advanced. In April, we showcased a preview where the robot’s punches and kicks were slow. By May, for a global CCTV broadcast, we doubled the speed and force, making it much more dynamic. This technology supports 15-20 random motion sequences for fighting, ensuring smooth transitions and resilience to external impacts, unlike dance performances. The aluminum alloy legs of our robots show dents from intense strikes, highlighting their durability.

The fighting experience is best for operators or those standing nearby, less so on video. Our new R1 humanoid robot, priced at 39,900 RMB, is lightweight (25 kg) and safe, with strong customization potential.

Let me share my thoughts on the robotics industry. This year, driven by high demand and policy support, robot and parts manufacturers have seen 50-100% growth—a rare phenomenon. Globally, companies like Tesla plan to mass-produce thousands of humanoid robots this year and release their third-generation model. Major players like NVIDIA, Apple, Meta, and Open AI are heavily invested in this field.

A common misconception is that hardware limits robot applications. In reality, current hardware, including dexterous hands, is sufficient. The real bottleneck is embodied AI, which isn’t advanced enough for large-scale humanoid robot applications. We’re at a stage similar to 1-3 years before ChatGPT’s breakthrough—everyone sees the direction, but no one has cracked it yet.

I believe the “ChatGPT moment” for robots will come when a robot can enter an unfamiliar room, take a bottle of water to someone, or tidy up autonomously after a simple command. This could happen in 1-3 years if fast, or 3-5 years at most.

The issue isn’t data but models. The industry focuses too much on data quality and not enough on model architecture, which isn’t unified or advanced enough. For example, Vision-Language-Action (VLA) models are simplistic and struggle with real-world interaction data. Adding AI training to VLA models isn’t enough; they need further optimization.

Recently, Google released a video generation model, essentially a world model, which we experimented with last year. We trained a model to generate a video of a robot performing a task, then used it to control a real robot. This approach might converge faster than VLA models, though it’s GPU-intensive. For robots, high-quality video generation isn’t necessary—just enough to drive actions.

Another challenge is reinforcement learning (RL) scaling laws in robotics. Unlike language models, where new training builds on prior knowledge, robot RL requires retraining from scratch for new tasks, which is inefficient. Solving this could lead to breakthroughs and great research papers.

In the next 2-5 years, end-to-end embodied AI models will be critical, alongside lower-cost, longer-lasting hardware. Producing millions or billions of humanoid robots involves immense engineering challenges, like those in the automotive industry.

Robots can’t carry large-scale computing power due to size and battery constraints—peak power is about 100W, equivalent to a few smartphones. Future solutions will rely on distributed computing, like factory-based servers for low-latency connections. In residential areas, localized computing clusters could serve households, reducing costs for new buyers.

The robotics and AI fields are a global co-creation effort. Chinese and U.S. companies, like NVIDIA, have made significant contributions, and no single company can dominate AI innovation. As Open AI and DeepSeek show, breakthroughs often come from young, innovative minds. The robotics industry thrives on global collaboration, and every company and university plays a role. Thank you all!

Exclusive interview with People’s Daily:

Since the beginning of this year, Unitree Robotics’ humanoid robots have frequently gained attention, and China’s humanoid robot industry has also received unprecedented focus. There have been praises and encouragement, as well as criticism and skepticism. Recently, at Unitree Robotics’ headquarters in Hangzhou, Zhejiang, our reporter conducted an exclusive interview with Wang Xingxing, the founder of Unitree Robotics, to hear this post-90s entrepreneur’s response to public attention and his thoughts on the industry’s development.

— Editor

“In the long run, this might just be a small spark, like the birth of the internet back then.”

Q: After appearing on the CCTV Spring Festival Gala stage, Unitree Robotics’ humanoid robots became extremely popular. Did you anticipate this?
A: Honestly, it was completely unexpected. In 2021, our “robotic cow” appeared on the Spring Festival Gala, and this year we initially expected a similar level of impact. But the result far exceeded our expectations—many people remembered the image of robots dancing. In a way, it became a landmark event combining traditional culture with high technology, a cultural symbol, and I feel very honored.

Q: With humanoid robots suddenly in the spotlight, how has this affected your company?
A: The high level of attention is good for the entire industry. Our company, other related companies in the industry, including parts suppliers, have all seen excellent performance in the first half of this year.
Of course, with more attention comes more pressure. We’ve had more visitors, inspections, and interviews at the company. I personally oversee technology and products, and these activities do take up some time. For the industry to grow, attention is necessary. There will be ups and downs—today the attention might be high, tomorrow it might fade, but ultimately, things will return to rationality.

Q: There are also skeptical voices amid the attention?
A: It’s normal for some to think it’s great and others to disagree. If a company or product is universally liked, that’s actually abnormal.

Q: How far are humanoid robots from being applied in daily life?
A: Humanoid robots entering everyday life and every household are still quite far off. The industry is still in its early stages. So, we’re starting with smaller applications, like scientific research, education, service displays, simple industrial uses, and as development platforms for AI companies.
Our biggest goal, and what we’re working on now, is to have robots perform tasks—various movements to help people with different jobs. But at the moment, the technology isn’t mature enough for widespread use in homes or factories. Steady progress each year is positive, and it helps foster greater understanding and tolerance for the industry.

Q: What’s the biggest challenge for large-scale adoption of humanoid robots?
A: The biggest challenge is that the AI level for robots isn’t advanced enough and hasn’t reached a critical breakthrough point. This is a universal challenge worldwide, and everyone is working on it. Breakthroughs in AI can happen unexpectedly—problems unsolvable today might suddenly be resolved at some point in the future.

Q: How long do you think it will take to break through the bottleneck for large-scale adoption?
A: Robot AI technology is progressing very quickly. For example, six months ago, I wouldn’t have believed our robots could compete in combat matches—I thought it would take one or two years, but we achieved it in just a few months. However, predicting industry development is hard. There might be noticeable progress in 3 to 5 years, but entering ordinary households will take longer due to ethical and safety concerns.
As long as AI technology keeps advancing, the robotics industry will grow stronger. This is a chain reaction driven by AI, and the trend has been clear over the past two to three years. People’s imagination for AI is expanding, and they’re becoming more optimistic about the robotics industry. The buzz around humanoid robots this year might just be a small spark in the long run, like the birth of the internet.

Q: What’s the significance of organizing robots to participate in various competitions at this stage?
A: As long as people enjoy it, providing emotional value is valuable in itself—it creates new demand. Robot performances and competitions are an important part of building the future robotics industry.

“We currently have relatively fewer AI talents, but in the long term, our competitiveness is very strong.”

Q: Unitree started with robotic dogs and has now shifted to the humanoid robot track in recent years. Why make this choice?
A: It’s mainly due to AI technology. In earlier years, AI was single-function, like recognizing numbers, images, or faces. In recent years, multimodal general-purpose AI has emerged, capable of many tasks, which requires a general-purpose robot, like the humanoid robots we’re developing now.

Q: Why humanoid robots specifically?
A: It’s simple—more people like humanoid robots than dislike them. That’s the main reason, and the market is very rational.

Q: Where does China’s humanoid robot industry stand globally?
A: We have an advantage in manufacturing, with high global shipment volumes. Every country has its strengths. We currently have fewer AI talents, but in the long term, our competitiveness is very strong. Additionally, with strong manufacturing and hardware capabilities, Chinese robotics companies are more cost-competitive and have lower barriers to use, which will benefit large-scale adoption in the future.

Q: How do you view competitors and their products?
A: The industry is very hot right now, with many companies involved, each with its own strengths. Maintaining an open mindset and healthy, reasonable competition is beneficial for the entire industry.

Q: How important is computing power to the humanoid robot industry?
A: It’s extremely important, especially large-scale cluster computing. Robots, particularly humanoid ones, rely on batteries, and battery load and power consumption limit their ability to carry large-scale computing power.

Q: How do you address the computing power issue?
A: I think there will be a lot of distributed computing in the future. If a factory has a centralized computing center, robots can connect to it. These centers are close by, so whether wired or wireless, the bandwidth is sufficient, and latency is low. But if a robot in Beijing connects to a computing center in Shanghai, the latency would be too high.

Q: Can the current talent supply meet the industry’s development needs?
A: The humanoid robot industry has always lacked talent. This industry didn’t exist before, and related talent has only been cultivated in recent years. Also, the industry is still niche, and the number of people entering it isn’t as large as people think. Both the quantity and quality of talent will take time to match industry growth.

Q: Which types of talent are most needed?
A: We’re short on all kinds of talent—technical, non-technical, business, HR, etc. The core shortage is still AI talent.

Q: Have you considered collaborating with universities or research institutes?
A: We’ve always had close ties with many universities domestically and internationally. Promoting technological progress in robotics and AI and building connections with universities and young people are very important.
Driving industry progress together has been our greatest value to the industry over the past few years. In a sense, a company’s performance is one thing, but contributing to the industry and society is even more important.

“We hope more people, especially young people, will join in and drive industry progress together.”

Q: Unitree published a book on quadruped robots and open-sourced its technology. Why do this?
A: Besides the book, we release open-source projects every few months. We want to promote industry development—when the industry thrives, companies within it can grow.
It’s also about cultivating talent. Whether through university collaborations, publishing books, or releasing open-source projects, we hope more people, especially young people, will join in and drive industry progress together.

Q: Why is it hard to cultivate R&D talent quickly?
A: The industry changes rapidly, so maintaining a learning habit is crucial. I do this myself—days are busy with company matters, but in the evenings, I have more personal time to read papers, write code, and keep learning and improving.

Q: How do you view the relationship between R&D and investment?
A: Whether learning new technologies or developing new products, the key is to clarify the direction. Most R&D investment goes into exploring directions, which is the most “expensive” part.
Staying highly sensitive to technology and products is crucial—being sharp enough reduces detours. Exploring with minimal cost to find truly worthwhile directions and then focusing investment there is key.

Q: What’s your ultimate vision for the robotics industry?
A: To truly drive technological progress and elevate human productivity to a new level. The invention of the steam engine and electricity transformed society’s productivity. Now, general AI and general-purpose robots could bring even greater changes, gradually freeing humans from strenuous manual labor.

Q: Unitree was born in Hangzhou. What’s the relationship between a city and tech companies?
A: Zhejiang has a very open and inclusive business environment. Tech companies’ growth is closely tied to young people mastering emerging technologies. Cities that attract young people will inevitably excel in emerging industries.

Q: What support does the robotics industry still need?
A: The country is already very supportive. I hope people will maintain an open mindset toward new technologies and products. Industry development cycles naturally have ups and downs—today, people might be optimistic, tomorrow less so. Staying open to these fluctuations and maintaining confidence in the future is the greatest support for this industry.

Interview with NetEase Technology

Q: Unitree Robotics has reduced the price of its robots to 99,000 RMB, and the newly released robot this year starts at 39,900 RMB. The price drop is quite rapid. You previously mentioned that robots might become free in the future. Is that achievable?
Wang Xingxing: We’re indeed lowering prices, though different models vary. Smaller robots are cheaper, while larger ones are more expensive. Fundamentally, we aim for better pricing so more people can afford and use them.
From last year to this year, we’ve seen a surge in customers purchasing humanoid robots, which has created a new ecosystem, including leasing markets and secondary developers. The key prerequisite is sufficient shipment volume.
Once robots can truly perform significant tasks, the global perception of robots will change dramatically. I even think that when robots can handle substantial work, governments might require companies to pay taxes per robot produced, based on the value they generate. For example, if a robot performs tasks, a portion of its output value could be taxed and paid to the state. Robots could even work autonomously—say, reclaiming wasteland, with part of their output going to the government.
But this depends on robots being able to perform human-like tasks. If things progress quickly, we might see significant changes in 2-3 years; if slower, maybe 3-5 years.

Q: What are Unitree Robotics’ plans for tackling AI technology challenges?
Wang Xingxing: AI progress in coding is very fast, and major companies are paying close attention. Simple programs or code have high success rates, but complex code sees a significant drop in success. The biggest technical challenge is improving embodied AI models, which are currently far from sufficient.
The goal is to reach a point where, in a venue like this, you can casually ask a robot to do something, and it gets it done. For example, if I can’t find food here and ask the robot to buy some, achieving that result is our target.

Q: What’s Unitree Robotics’ stance on AI investment?
Wang Xingxing: We remain cautious. Our AI team is relatively large within the company, but compared to major AI firms, we can’t compete in scale.
However, I believe there’s opportunity. In AI, success often isn’t about having the most resources, money, or people—it’s about the best solution emerging globally. Maintaining sufficient investment, like having a small or medium-sized team produce better results, has a decent probability of success.

Q: How will Unitree address the lack of generalization and autonomous decision-making in robots?
Wang Xingxing: This can be boiled down to one issue. Current hardware is sufficient, though not perfect—larger scale, lower cost, and higher reliability need improvement. The biggest issue is that AI models haven’t reached a breakthrough critical point.
A ChatGPT-like moment for embodied AI hasn’t happened yet.

Q: This year, humanoid robots have participated in many sports events, and people are curious about your personal interests. What new progress has Unitree made in sports-related applications?
Wang Xingxing: I don’t exercise much personally. I used to cycle more, but work keeps me busy now.
From last year to this year, our company has focused more on performances like dancing or fighting competitions. Next week, there’s a robot sports event in Beijing—everyone can check it out.
Current technology needs time to mature. Expecting a humanoid robot to work in homes or factories with significant practical value right now is unrealistic. Demos are fine, though.
At this stage, using AI for sports and performances is an easier scenario to implement. Early computers needed developers to create software before they became useful to the public—it’s a similar phased process.

Q: What’s Unitree Robotics’ product strategy and approach to commercialization?
Wang Xingxing: Our humanoid robots have seen rapid sales growth from last year to this year. Most of the robots our customers use are work-oriented, so we have a significant share of the global humanoid robot market.
From a size perspective, larger robots, around 1.7 meters, are better suited for industrial tasks or practical work. Smaller robots, like 1.2 meters, are aimed at developers or entertainment purposes. Expecting a 1.2-meter robot to do heavy work is too challenging, so we’ve lowered the price threshold for those.

Q: Unitree has humanoid and quadruped robots, as well as whole systems and modular components. How do you coordinate so many product lines?
Wang Xingxing: We’re hiring more people each year, which raises management demands. Otherwise, more staff could lead to lower efficiency.
Our products share many components, which helps. Effective management is critical for a company’s growth.

Q: How do you view the current state and future development of the robotics industry?
Wang Xingxing: Our company’s goal, and the industry’s overall goal, is to have robots genuinely work and liberate productivity. That’s the most critical gap right now. However, objectively, people prefer performance displays over work demonstrations.

Q: What’s your take on current robot shipment volumes?
Wang Xingxing: The industry can sustain a doubling of humanoid robot shipments annually. With a technological breakthrough, we could see hundreds of thousands or even millions of units shipped in a single year within the next 2-3 years.

Q: What’s the timeline for Unitree Robotics going public?
Wang Xingxing: We’re steadily working toward a standard IPO process. Having been established for nine years, I view going public as a learning and growth process, like preparing for a college entrance exam. It’s a milestone for more mature management and operations, a reflection of our past nine years, and an accountability to our shareholders.