Why Unitree's Wang Xingxing publicly admitted his mistake: general-purpose large models disrupt the embodied intelligence R&D path

Wang Xingxing admitted he was wrong. In the October that is about to pass, his worries about the future seemed to suddenly accelerate.

At the Industry Expo on October 12, Wang Xingxing said, "The biggest mistake I have made over the past decade or so is that I did not believe in AI enough." Ten days later, at the Physical AI Conference, his wording was even stronger, saying bluntly, "In recent years, my expectations for AI could still be more than 10 times higher."

That Wang Xingxing would publicly admit a mistake is especially important for the embodied intelligence industry.

After all, he is the person who turned Unitree Technology into the "first listed humanoid robot company." For quite a long period, Wang Xingxing treated "a robot is first of all a machine" as a default premise. Moreover, Unitree's revenue in the first three quarters of 2026 was 1.817 billion yuan, up 55.71% year on year, and the G1 also became one of the best-selling humanoid robots in the world.

His approach seemed to be successful in the past.

However, the embodied intelligence industry is currently undergoing a subtle split. On the one hand, the deployment of embodied intelligence is still stumbling. Wang Xingxing admitted that the generalization and efficiency of large-scale robot applications "still need to be further improved." On the other hand, general large models are refreshing the boundaries of capability at an almost unreasonable speed. Especially after the release of GPT-6 Astra, Unitree connected it to its own robots for testing, and Wang Xingxing said "the effect is actually very good."

The reaction of the capital market is also interesting.

On October 15, Unitree Technology fell to a low of 404 yuan, and its market value evaporated by more than 240 billion yuan from the high on its first day of listing. However, as Wang Xingxing began to "admit mistakes" and began to say that "physical AI has been breached," the company's stock price launched a desperate counterattack, rebounding from 404 yuan to once exceed 700 yuan, rising 70% in less than two weeks. On October 25, JPMorgan directly raised Unitree's rating by two notches from "underweight" to "overweight," and adjusted the target price from 300 yuan to 995 yuan. The reason was very direct: Wang Xingxing compressed the expected breakthrough in physical AI from "at most 10 years" to "2 to 3 years," and judged that the industry's sales S-curve will clearly move forward.

So, behind Wang Xingxing's "admission of mistakes" and worries, what direction for embodied intelligence is actually being foreshadowed?

Why did Wang Xingxing admit he was wrong?

To understand Wang Xingxing's "admission of mistakes," one must first understand how embodied intelligence worked in the past.

This path is simple and basically fixed. It can be divided into several processes: defining tasks, breaking down actions, writing control algorithms, and repeatedly debugging in specific scenarios. Unitree's G1 can run, jump, do backflips, and even twirl handkerchiefs and do the yangge dance on the Spring Festival Gala stage, relying on engineering accumulation at the motion control level.

This method works very well in environments with clear boundaries. A well-tuned robot can repeat thousands of times at a fixed workstation without error. However, the ceiling of this kind of robot that requires pre-training is also obvious. Change to another scenario, another object, or another task, and all the previous rules may become invalid, and training has to start over from scratch.

Robots are different from large language models. Sensors have errors, actuators have deviations, and when grasping unfamiliar objects, a difference of a few millimeters can cause the success rate to plummet. In such an environment, traditional engineering thinking prevails, because errors can be measured, deviations can be calibrated, and problems can be solved one by one.

However, after the emergence of GPT-6 Astra, the impact it brought may overturn the traditional path. And in tests conducted by Unitree Technology itself, what deeply surprised Wang Xingxing was not that the robot could draw, but the condition of "zero training."

In the past, for a robot to learn a new task was quite laborious. It required collecting data, training models, and tuning parameters. The whole process took at least several weeks and at most several months. GPT-6 Astra completely skipped this process. Giving it a prompt was enough, and the robot could begin acting according to the requirements.

This surprised Wang Xingxing very much, and he also felt pressure. Otherwise, he would not have publicly admitted a mistake.

This process of GPT-6 Astra means that the "real-machine data accumulation" and "task training pipeline" that the embodied intelligence industry was once proud of are, at least for some tasks, no longer indispensable processes in the face of general large models.

While you spend months painstakingly collecting data and training models, an external team using a large model can complete the same task on the same body with zero training.

Months of training versus zero training—such a contrast leaves practitioners no room not to worry.

Wang Xingxing's "admission of mistakes" is not a simple piece of self-criticism. What he truly admits is that the problem-solving approach long taken for granted by the embodied intelligence industry may not be the only path, and may even not be the fastest path. The capability growth of general large models may be overturning the foundation of the entire industry.

Of course, this does not mean that traditional robot technology will be eliminated.

According to Tianyancha, Unitree Technology's intellectual property accumulation is mainly concentrated at the hardware and control levels. It currently has more than 100 robot patents, including motion control methods for humanoid robots, boxing control methods and systems, and so on. Unitree's motion control, hardware design, and mass production capabilities are still solid. But such technology may no longer be enough to "constitute a moat."

So, Wang Xingxing admitted he was wrong. But even if large models have achieved rapid advances, will the bottlenecks of embodied intelligence disappear on their own?

How can the bottlenecks of embodied intelligence be broken through?

At the Industry Expo, Wang Xingxing brought the narrative back to reality.

"When GPT-7 comes out next year, it may directly break through the embodied intelligence problem." That statement sounds like he has been won over by large models, but in August and October, Wang Xingxing also repeatedly emphasized another line: "The core challenge for embodied intelligence to pass the ChatGPT moment is still the alignment between the input and output of AI models and real robots. Large models can plan and navigate, but what happens after physical contact has not yet been solved."

There is still a huge distance between "being able to talk" and "being able to do."

For large language models, the input is text and the output is also text. If it makes a mistake, it can be withdrawn, and the cost of starting over is almost zero. But a robot situated in physical reality is completely different: the input is sensor signals, and the output is joint torque. In this process, sensors have noise, actuators have latency, objects have friction coefficients, and the environment has lighting variations. Subtle changes can all lead to completely different results.

Wang Xingxing gave a specific example: asking a robot to pick something up. It looks like it is about to grasp it, but in the end it is off by a few centimeters or even a few millimeters, and the error cannot be corrected, so the success rate drops severely. For this kind of problem, although GPT-6 can understand the instruction "throw the paper box into the trash can," it cannot understand the change in the center of gravity after the paper box is crushed flat, nor does it understand the rebound force at the edge of the trash can or the effect of a slippery floor on gait.

These problems cannot be solved by a larger number of parameters or more text training data. Because the laws of the physical world require data from physical interaction to learn, and the collection of real-machine data is precisely slow and expensive, making it difficult to scale.

Simply put, the idea is to let AI not only control robots, but also accelerate the R&D process of robot capabilities. AI retrieves papers itself, writes control code, trial-and-errors in simulation, deploys to real machines for testing, and iterates based on the results. If this closed loop can run, the efficiency of data collection and model iteration will increase substantially.

However, this approach still has a prerequisite: the gap between the simulation environment and the real world must be small enough that what is learned in simulation does not completely fail on real machines. And this "small enough" has not yet been proven.

Now it seems that Wang Xingxing's position looks somewhat contradictory. He says both "break through within one or two years" and "the alignment problem is not solved." In fact, from the growth curve of large model capabilities, the trend is very steep, while from the reality of physical execution, every step is stumbling. Both are real.

Wang Xingxing can neither pretend not to see the development of large models nor directly pass through the bottleneck of embodied intelligence. Therefore, his "admission of error" acknowledges the speed of the former, while his "worry" is that the technical bottleneck will not be solved just because the former becomes faster.

This contradiction can be glimpsed in Unitree Technology's financial report data. In the third quarter, its revenue was 665 million yuan, up 69.93% year on year, but net profit attributable to shareholders was 113 million yuan, down 17.37% year on year, and net profit after deducting non-recurring gains and losses fell 17.42%. Where did all the money go? R&D.

Currently, the proportion of Unitree Technology's R&D investment in revenue has jumped directly from 9% in the same period last year to more than 19%. Wang Xingxing does not dare to stop. Before the bottleneck of physical execution is broken through by AI large models, he wants to strengthen his own capabilities as much as possible.

His anxiety is much more complex than "whether robots can be built."

Unitree Technology's pressure and way out

In fact, on the embodied intelligence track, Unitree Technology's robots are doing quite well. Its robots have appeared on the Spring Festival Gala for consecutive years, cumulative sales have reached tens of thousands of units, and it can still remain profitable while peers generally lose money.

But this is not the key to the problem. The situation Wang Xingxing needs to face is not optimistic, because when the way to solve "making robots work" changes, whether the entire set of advantages Unitree has built around "building robots" can still be converted into future competitiveness has become a question. The shadow of a "dimensionality reduction strike" is already hovering overhead.

There are threefold pressures here. The first is to redefine R&D speed. At the Industry Expo, Wang Xingxing described the current state of robot model development: researchers have ideas, search papers, write code, train models, test, and provide manual feedback. One round of iteration takes a very long time.

According to his latest judgment, AI can automatically complete these steps. If that holds, the gap between companies will no longer depend on "how many engineers each has," but on "whose AI closed loop runs faster." At the same time, the surge in R&D investment and the financial data showing revenue growth without profit growth also indicate that this closed loop is still in the "burning money" stage and may not immediately translate into profit.

Spending money to buy speed is one thing; whether speed can be bought is another.

The second pressure is the migration of technical barriers. To put it simply, Unitree Technology's past barrier was that it could make robots cheaper and make them able to run and jump. But when general large models make "making robots work" cheaper and more general, the advantage in hardware cost will be compressed into manufacturing profit.

Wang Xingxing himself listed a checklist, including large-scale manufacturing capability, proprietary physical AI models, ultra-large-scale computing clusters, chip supply, and energy supply. Among these, Unitree has accumulated experience in manufacturing but does not have an advantage in models and computing power. He also admitted that it is "not very realistic" for Unitree to "build exactly the same models as the world's top multimodal large model companies."

This is where he is clear-headed, and also where he is helpless.

The third pressure, which is also easily masked by the stock price rebound, is the time lag between technological breakthroughs and commercial success. Wang Xingxing believes that after GPT-7 is released, it may break through embodied intelligence. However, "breaking through" and "making money" are two different things. For physical AI technology to land at scale, it still faces multiple challenges such as perception accuracy, motion control stability, data accumulation, and software-hardware coordination.

Unitree's choice is already clear: on the one hand, continue making hardware, push shipment volume upward, and make scale itself a bargaining chip. On the other hand, bet on "physical AI self-evolution," betting that before general large models completely erase the premium of "body hardware," it can gain control over the party that has deployment scale and data feedback in real scenarios, thereby gaining pricing power.

Wang Xingxing said, "I am really very afraid of wasting every day."

His sense of urgency is not fear of any one company, but a sense that a window is narrowing and there is no time to lose.

Perhaps the stock price rebound gives Wang Xingxing more time and space to make another gamble. But it also pushes the cost higher. And what he truly needs to be vigilant about may be that Unitree's past path to success may not necessarily allow Unitree to live well tomorrow.

This article comes from the WeChat public account "Rhodium Technology," author: Hongchen, published by 36Kr with authorization.