Wang Xingxing's Speech at Unitree's 10th Anniversary: Urging People to Believe in AI, the Embodied Intelligence Window Narrows to 1 to 2 Years

On October 22, the Hangzhou International Expo Center was buzzing with activity—not only were there many people, but there were also especially many robots.

This was the scene of the 2026 World Physical AI Conference hosted by Unitree Robotics. From quadruped robots to humanoid robots, you could even say this is the place with the highest "robot density" in the world.

2026 also marks the tenth year since Unitree Robotics was founded. Interestingly, on today's stage, Unitree founder Wang Xingxing, aside from talking about robots, spoke most about the hottest topic right now: AI.

01. You Must Believe in AI

In Wang Xingxing's 36-minute speech, Unitree's own products only occupied the beginning and the end. The longest section in the middle was Wang Xingxing doing one thing—persuading people to believe in AI.

From AI conquering a batch of top-level math problems—"the tip of the pyramid of humanity's greatest intelligence" has been breached; to AI reverse engineering making "the moats of traditional software companies disappear"; to the expressions and emotions of characters in AI videos being "much better than what real people act"; to top large-model companies having "already achieved self-evolution of models" internally.

"This is not a joke." His conclusion was that with current model capabilities, even if they stop improving, the information world has already been "completely breached."

Then he said something that didn't quite sound like something a listed company chairman would say: "I feel that I myself still don't believe enough, still aren't firm enough."

This was not improvisation. Ten days earlier, at the Shanghai Industrial Expo, he had already "admitted his mistake" in public once: "The biggest mistake I've made over the past decade or so is that I didn't believe in AI enough." He also gave a multiplier: the degree to which you liked and believed in AI 5 or 10 years ago—"multiplying it by 10 now would not be an exaggeration."

The more intuitive change is in the timeline. At the World Robot Conference in August this year, Wang Xingxing's expectation for the "ChatGPT moment" of embodied intelligence was still 2 to 3 years at the fastest, or 5 to 10 years at the slowest; two months later, he compressed that window to "1 to 2 years, or 2 to 3 years if slow." As for casually bringing a Unitree robot into an unfamiliar home and telling it in one sentence to tidy up the room, "at the fastest it might be achievable next year." At the Industrial Expo he put it even more bluntly: "When GPT-7 comes out next year, it may directly break through the embodied intelligence problem."

But Unitree's strength—the robot body itself—remains something Wang Xingxing is proud of.

What Wang Xingxing was most proud of on stage was the G1. This humanoid robot, released in May 2024, has cumulatively rolled off the line about 10,000 to 20,000 units, which he called the humanoid robot with the largest shipment volume for a single model globally.

On October 21, Unitree released the H2-X "Kung Fu Legend: Guardian" video: kicking a car door in, lifting a car bare-handed, landing a backflip from a 4-meter-high platform, with battery life extended from about 3 hours on the H2 to about 8 hours. It can be seen that the Unitree team is still rapidly iterating the company's products.

He was very candid on stage: "After all, we are a new company," and "making an identical model" to the world's top multimodal large-model companies "is not very realistic." Unitree will develop its own models and also cooperate with large-model companies at home and abroad. The most eye-catching demo he showed was connecting GPT-6 to a robot, which, with zero training, could wash brushes, change paint, and precisely place pieces on a Go board by itself. He judged that problems like planning and navigation "have, in a sense, already been solved," and what remains is the alignment between the input and output of multimodal large models and real robots—sensor errors, actuator errors—"I think this is also something that can be solved very quickly."

Connecting these few sentences makes clear the business logic behind this "faith." If the robot's brain is ultimately provided by companies like OpenAI and Google, and, as he says, becomes increasingly powerful and increasingly cheap, then the deciding factor in the embodied intelligence race will return to who can build reliable bodies most cheaply and at the largest scale. This is precisely what Unitree is best at.

So among the most important things Wang Xingxing listed for the future, the first is not models, but large-scale hardware manufacturing capability. Unitree currently produces tens of thousands of robots and hundreds of thousands of joint motors a year, which in his view is "completely insufficient"; true popularization requires at least "annual production of ten million units." He even talked about computing clusters, chip production capacity, and mines. As for Unitree's own model capabilities, what he is betting on is a "self-evolution" loop that lets AI retrieve papers on its own, write control code, train in simulation, test on real machines, and then feed the scoring results back. The premise of this approach is likewise that foundation models get stronger every month; in his words, "the more model capabilities improve," the more this loop "will also become stronger and stronger."

At the end of the speech, Wang Xingxing talked about a "crossing timeline" between himself and AI. In 2009, as a college freshman, the first robot he made was a bipedal humanoid; for a computer course paper in college, his topic was "as the scale of neural networks increases, capabilities will become stronger and stronger," on the grounds that the human brain is larger than those of chickens, dogs, and pigs, so humans are smarter. Then came AlphaGo in 2016, and ChatGPT at the end of 2022, the latter of which directly prompted Unitree to formally enter humanoid robots in 2023.

This recollection is told very sincerely, and also very selectively; it turns more than a decade of a hardware engineer's life into a straight line that always pointed toward AI. Maybe he really did believe early on, only he didn't believe enough; or maybe this era forces every robotics company to tell the same story.

02. Beyond the bustle of the booths

At this World Physical AI Conference hosted by Unitree, the exhibition area could be described as the classic scene of more robots than people. Besides traditional showcases such as ring fights and table tennis, you could even see robots doing diabolo and playing "kendama"—yes, the developer of the kendama-playing robot was speaking Japanese.

The kendama-playing robot | Image source: Geek Park

Robot dogs of all shapes and forms | Image source: Geek Park

Across the entire exhibition area, we could see that at present, whether quadruped or humanoid robots, the biggest scenario is probably still guided tours and entertainment. For example, you could see a G1 robot wearing a cowboy hat, and also quadruped robot dogs dressed up as pandas, dinosaurs, and lions for lion dance. A company called Yijing Technology even equipped its robots with "faces" that can make expressions. The team lead told Geek Park that in scenarios such as eldercare and guided tours, a robot with a "face" still feels more approachable than an ordinary robot, despite a bit of the "uncanny valley" effect.

In addition, scenarios such as security and patrols are also truly being implemented. The lead of a company that modified robot dogs for patrol scenarios told Geek Park that quadruped robot dogs equipped with lidar and related equipment can already replace humans on night duty in residential compounds and industrial parks, and in critical situations, human operators can also take remote control. And if equipped with firefighting equipment on their backs, these quadruped robots can even carry out tasks such as firefighting.

A staff member is operating remotely | Image source: Geek Park

A relatively popular booth in the exhibition area was the one for robot education and courses. A Unitree staffer told Geek Park that the team has a mature solution that can provide not only hardware, but also a package covering hardware, software, and talent, offering schools a complete robot education laboratory so that more students can truly get hands-on experience with the appeal of robot programming.

A robot learning to shoot hoops | Image source: Geek Park

At the conference itself, Wang Xingxing and a group of guests also launched the World Physical AI Competition, with Unitree offering prizes worth tens of millions of RMB to encourage developers around the world to develop robots. This is also a rare domestic competition that truly cultivates robot developers since DJI's former "RoboMaster" competition.

Just as the new wave of AI in recent years has cultivated countless AI entrepreneurs and super-individuals, when physical AI begins to enter a critical state, this industry likewise needs countless robotics talents, and the World Physical AI Competition is undoubtedly a very good starting point.

*Header image source: Unitree Technology

This article comes from the WeChat public account "Geek Park" (ID: geekpark), author: Jing Yu, published by 36Kr with authorization.