Unitree's Wang Xingxing: GPT-6 Astra Impacts Embodied Intelligence, Physical AI May Break Through in One to Three Years

On October 22, 2026, Unitree held the "2026 World Physical AI Conference and Unitree Partner Ecosystem Conference" in Hangzhou, marking the first time in Unitree's ten-year history that it has held an ecosystem partner conference.

"Almost everyone in the world, including myself, still does not place enough importance on and faith in AI," Wang Xingxing said. In his view, even though AI has become the world's most closely watched technology direction, people still underestimate its pace of development and potential impact.

He gave an interesting example. In fact, four or five years ago, he had recommended some AI-related stocks to company colleagues. "If we had believed in AI more a few years ago, the personal wealth of some Unitree employees might already have grown tenfold or even dozens of times."

Wang Xingxing speaking

He took the internet as an example: from the dot-com bubble around 2000 to the gradual popularization of applications such as e-commerce around 2010, the whole process took nearly a decade. Therefore, for the most advanced AI models to truly enter ordinary people's lives, time is likewise needed for fermentation.

As the founder of a robotics company, he believes the opportunity lies in physical AI. Wang Xingxing even offered a fairly aggressive judgment: as the capabilities of general large models improve rapidly, physical AI may see a major breakthrough in the next one to three years. "One to two years if fast, two to three years if slow."

Unitree displayed the newly released H2-X robot

GPT-6 Astra deals a huge blow to embodied intelligence

Wang Xingxing believes that the recently released GPT-6 Astra has dealt a huge blow to the industry. Inside Unitree, they also tried using this model to conduct several robot experiments.

One of them was having a robot paint. According to his account, Unitree had not written complex programs for this painting task in advance; they simply used prompts to have the robot complete the painting. Surprisingly, the robot also autonomously completed supporting painting tasks such as changing paint and cleaning brushes.

Another experiment Unitree conducted internally with GPT-6 Astra was playing Go.

They found that the robot could not only recognize the board but also precisely place the stones in the corresponding positions. Wang Xingxing repeatedly emphasized in his speech that these experiments were all completed under "zero-training" conditions.

In the past, robots often needed programs written for specific tasks or specialized training. Now GPT-6 Astra has proven that general large models already have the potential to directly control robots to complete some tasks.

However, Wang Xingxing also admitted that GPT-6 Astra is not omnipotent. In his view, large models can already solve robot task planning and navigation problems fairly well, but when it comes to actual manipulation, they are still prone to errors.

For example, getting a robot to pick up an object is relatively easy, but getting it to assemble precision components is much more difficult—because the robot must not only know what to do but also adjust its movements in real time based on the object's position, force conditions, and so on. Wang Xingxing believes that current large models still have obvious shortcomings in perceiving changes in the real environment and correcting operational errors in a timely manner.

Wang Xingxing judges that the core challenge for whether embodied intelligence can enter its ChatGPT moment next is whether the input and output of AI models can be aligned with real robots.

Wang Xingxing speaking

He also announced his vision. He hopes that by the end of this year or next year, Unitree robots will be able to enter ordinary households, where people can use voice commands to have robots complete tasks such as tidying rooms. The longer-term goal is that robots may be able to smoothly complete 80% of tasks through verbal instructions.

On the brain side, Wang Xingxing said that in addition to continuing to advance cooperation with large-model companies at home and abroad, Unitree is also simultaneously advancing its own physical AI system. Currently, Unitree is developing a model system called "Physical AI Robot Self-Evolution V1."

In the past, each time a robot mastered a new skill, researchers often had to invest a great deal of time in data collection, training, and optimization. Even if success was achieved in the laboratory, transferring it to a new environment might require readjustment.

How Unitree's self-evolution system works

Wang Xingxing also explained why Unitree is developing this system.

In his view, the stronger the foundation model, the stronger the robot's self-evolution capability. As more data is used for training, the scale effect of self-evolution will gradually emerge. And when more robots are actually deployed, they will bring more real-world data, further driving model iteration.

Through this cycle, robots can continuously accumulate and expand skills, ultimately forming a scaled skills ecosystem.

Wang Xingxing explains why they are pursuing self-evolution

However, what Unitree has disclosed so far is mainly the technical approach. The actual training efficiency of this system, its skill transfer capability, and its success rate in real environments still remain to be verified.

In the past, Unitree's advantages were mainly concentrated in the robot body, including joint modules, power systems, and motion control algorithms. But as industry competition shifts from "making robots move" to "making robots complete actual tasks," hardware advantages alone are no longer sufficient to determine a company's long-term competitiveness.

Unitree has clearly recognized this change.

Wang Xingxing revealed that since last year, Unitree has begun using pretrained deep reinforcement learning technology and independently developed a pretrained whole-body motion control model, enabling robots to be trained in batches on different movements and greatly improving training efficiency.

Unitree's progress in models and data since 2026

This is the best of times, and I am very afraid of wasting every day

Wang Xingxing expressed optimism about the industry's future, but he cannot avoid the brutality of competition.

“This industry is extremely hot, the pie in the future is very big, the prospects are very broad, but the competition is also extremely fierce,” he said.

This sense of urgency is also reflected in his personal state. He said frankly that he has many things he wants to do every day; besides robots, there are many new technologies he needs to learn and explore, but time is increasingly running short.

“Now is the best of times, and I am very afraid of wasting every day.”

In his view, the foundation accumulated over thousands of years of human civilization has given this generation of humanity the opportunity to truly realize superintelligence or AGI, and to raise productivity to a new level through humanoid robots.

He even painted a picture of a more distant society. When robots take on a large amount of heavy labor and production, in the future, daily necessities including housing may, through a social distribution system, benefit more ordinary people.

On the hardware manufacturing side, he revealed that Unitree's G1 robot has already rolled off the line a cumulative 10,000 to 20,000 units. But this production capacity is still far from enough; he believes that in the future the robotics industry will need manufacturing capacity of 10 million units per year.

And as model capabilities continue to improve, the demand for robot training and computing power will also continue to grow. Ultra-large-scale server clusters, as well as manufacturing capacity for computing power and storage chips, will all become important foundations supporting the industry's development.

Based on this judgment, Wang Xingxing believes that the core resources of the future world will be five—data, computing power, chip production capacity, energy, and minerals.

Unitree's progress in models and data since 2026

Unitree doesn't just want to perform; it also wants to build an ecosystem

Over the past few years, Unitree has frequently gone viral with its robots' athletic abilities such as dancing, fighting, and running, but the outside world has always had a question: why are Unitree's robots always performing rather than doing work?

At this conference, Wang Xingxing also held a roundtable conversation with Yao Ming and responded to this question.

In his view, robots boxing and playing ball are similar to humans doing physical exercise; the purpose is to improve athletic ability, stability, and movement precision, laying a foundation for performing actual tasks in the future.

“If you can't even walk well or run, how can you expect such a robot to do work?”

Yao Ming in conversation with Wang Xingxing

Taking robot fighting as an example, he explained why Unitree is so keen on these sports that seem unrelated to practical applications.

On the one hand, fighting involves extreme situations such as collisions and falls, which can test the robot's reliability. On the other hand, fighting requires completing a large number of complex movements, placing very high demands on the robot's motion control algorithms and stability.

Beyond performances, Unitree is also trying to cooperate with more partners to push robots into real application scenarios. However, Unitree's strategy is not to enter the field itself, but to build an ecosystem.

At this ecosystem conference, Unitree officially launched the “Yuchuang Plan,” recruiting entrepreneurs in areas such as robot bodies and core components, embodied intelligence algorithms, data and simulation platforms, and industry applications and solutions, targeting teams that have already established companies or have clear entrepreneurial directions.

Wang Xingxing said it will provide support in industrial capital, robot platforms, development tools, engineering mass production, and application scenarios.

Yuchuang Plan

Wang Xingxing proposed that in the next five years, Unitree hopes to incubate more than 200 high-quality ecosystem enterprises through this plan.

This also means that Unitree is trying to expand from simply selling robot products to an ecosystem model of providing technology platforms and jointly developing applications with partners.

However, whether this can attract enough developers and startup companies, and turn these application explorations into sustainable commercial revenue, is still a question Unitree needs to answer.

This ecosystem conference is Unitree's first formal address to the industry after going public. As a benchmark company with a market value of over 100 billion, Unitree carries complex emotions. Today, it is no longer the private possession of that geeky young man; as one of the few leaders in the industry, it must lead a more exciting and grand story.