Unitree's Wang Xingxing: Physical AI May Reach a Tipping Point Within One to Two Years; GPT-6 Has Been Deployed to Robots with Zero Training

Smart East (Zhidongxi) reported from Hangzhou on October 22. Today, at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Wang Xingxing, Chairman, Founder and CEO of Unitree Robotics, delivered a roughly 30-minute speech, touching on important topics such as AI faith, the critical point of physical AI, the ChatGPT moment of embodied intelligence, self-evolution V1.0, and the mass-production ecosystem.

"Almost everyone in the world still does not place enough importance on and faith in AI, including myself," Wang Xingxing said. Physical AI has already shown signs of being broken through, and it is very likely to happen in the next 1 to 2 years; the industry is moving toward the "ChatGPT moment of embodied intelligence," in which "in 80% of unfamiliar scenarios, robots can complete about 80% of tasks according to voice or text instructions."

Wang Jian, Director of Zhijiang Laboratory and founder of Alibaba Cloud, also took the stage as a heavyweight guest to share his views. He proposed that the concept of AI has been somewhat overused and that it actually brings together many different technologies. He believes that AI and robots are very likely to converge and become something new, and that "soul" and "body" truly have a chance to truly become one.

In his view, large language models still leave huge room for imagination. Language cannot be separated from muscle actions such as reading, typing, and speaking. Incorporating information from these actions into models is also an opportunity for startups. He admitted that an executive of Boston Dynamics praised Unitree in the early years, but he himself failed to notice the company earlier, and said he "looks forward to the arrival of the AI Unitree era."

Unitree's new humanoid robot "H2-X Kung Fu Legend Guardian," released online yesterday, made its first appearance today. The company says it has a maximum battery life of about 8 hours, can lift a car with its bare hands, land smoothly from a height of 4 meters, and kick a car door dented with one kick.

At the event, Wang Xingxing also explained the pre-researched Physical AI Robot Self-Evolution V1.0, which relies on stronger foundation models, multi-source data, more real-machine deployments, and skill ecosystem expansion to form a cycle of perception, decision-making, execution, evaluation, learning, and evolution.

He envisioned that he hopes that from the end of this year to next year, if a user tells the robot to tidy up a room, it will be able to do so.

At the same time, a large number of partner solutions were showcased in a concentrated manner at the conference. After viewing them, Wang Xingxing remarked: "Even I have never seen such a comprehensive set of solutions." Unitree's platform-based ecosystem intention also became clearer accordingly.

He mentioned an anecdote on site: four or five years ago, he asked colleagues at the company to pay attention to the stocks of leading AI companies. If everyone had believed it, their financial situation might now have risen dozens of times. Now, there are still very few people who have truly seen the effects of top-tier AI capabilities.

Recently, AI has achieved more milestones, including conquering frontier mathematical research problems, software reverse engineering and vulnerability discovery and exploitation, high-fidelity video generation, robotaxis, and more; inside top large-model companies, model self-evolution has basically been achieved, and the iteration of model capabilities will continue to accelerate in the next one or two years.

So what about the physical world?

Wang Xingxing said that physical AI has now shown signs of being broken through, and it is very likely to happen in the next 1-2 years or 2-3 years.

He revealed that GPT-6 has now achieved completely zero-training deployment onto Unitree robots, and can autonomously paint, clean brushes, autonomously play Go and chess on a fine chessboard, autonomously find a kitchen and look for tools, and perform other tasks.

"We are moving toward the ChatGPT moment of embodied intelligence. In 80% of unfamiliar scenarios, through voice or text instructions, robots can smoothly complete about 80% of tasks."

Wang Xingxing mentioned a specific scenario: "I hope that from the end of this year to next year, you tell the robot to help me tidy up the room, and it can do it."

Second, robots must self-evolve: align errors and also compete in mass production of tens of millions of units

What is the core challenge for embodied intelligence to pass the ChatGPT moment?

When physical AI also breaks through the critical point, the most important points may include the following aspects: first, the ability to mass-produce robot hardware, with an annual production of tens of millions of units; second, self-developed or cooperative physical AI models and their implementation; third, ultra-large-scale computing power servers; fourth, computing power and storage chip manufacturing capabilities.

Wang Xingxing revealed that one of the most important things Unitree has done this year is the pre-research of Physical AI Robot Self-Evolution V1.0. As shown in the figure below, driven by the world's top AI large models, Unitree will step in to define skills, tools, task constraints, and more, promoting robots to achieve a self-evolution cycle of perception, decision-making, execution, evaluation, learning, and evolution.

To realize Physical AI Robot Self-Evolution V1.0, the following conditions are needed:

1. The stronger the foundation model, the stronger the self-evolution performance.

2. Large-scale multi-source data drives the scale effect of self-evolution.

3. More real-machine deployments, more data accumulation, and exponentially growing self-evolution speed.

4. Cumulative expansion of skills, moving toward a large-scale skill ecosystem.

Finally, Wang Xingxing also announced the timeline of his own interweaving with physical AI. He said that the present is the best era, and he is afraid of wasting every day. He always feels that there is still a lot to learn, but there is too little time. At the same time, competition is fierce, and one needs to do better every day.

Third, about 20,000 G1 units have rolled off the line, and global developers are putting robots into real scenarios

At the event, Wang Xingxing also reviewed Unitree's important milestones over the past year, from the 2026 Year of the Horse Spring Festival Gala "Wu Bot" to the world's first mass-produced manned mecha GD01, and then to the release of Unitree's general humanoid foundation model UnifoLM-WLA-1.0. Unitree's software and hardware have both developed fully.

Large-scale mass production provides the foundation for Unitree's broader exploration. Wang Xingxing announced a set of data: since its release in May 2024, the Unitree G1 humanoid robot has become the world's best-selling humanoid robot, with cumulative production of about 10,000 to 20,000 units.

Behind this, partners from around the world are accelerating the development of intelligent robots. A large number of open-source algorithms have been implemented by developers on the Unitree robot platform, covering many fields such as education and scientific research, industrial manufacturing, entertainment performances, embodied guided tours, service retail, intelligent inspection, and emergency firefighting.

Many developers from various parts of the world came to the conference site today. What impressed Smart East (Zhidongxi) deeply was Jake Laser, an American tech content creator and YouTube blogger.

He shared that based on Unitree's Unitree B2 quadruped robot product, he developed a "smart wheelchair" for his father, who has difficulty walking, ultimately successfully enabling convenient actions such as going up and down stairs, traveling, hiking, and wading through water.

Many partners' solutions were displayed on site. The ecosystem partner wall displayed a total of about 180 partners, covering fields ranging from computing power, the internet, industrial components, robots, and AI companies to vertical industries, universities, and research institutions.

2026 is a key year for the large-scale rollout of embodied intelligence. Data from the Ministry of Industry and Information Technology shows that in the first half of the year, China's humanoid robot shipments exceeded 40,000 units. However, the industry as a whole is still in its early stage, and the focus of competition is shifting from "body form" to intelligent generalization capability and ecosystem prosperity.

For Unitree, the convening of this ecosystem conference is a key declaration of its effort to build a "platform-based ecosystem company." Through forums and displays covering all dimensions such as channels, technology, talent, and capital, Unitree aims to seize the industry discourse and definition power of physical AI, accelerate industrial chain collaboration, and build competitive barriers through the ecosystem.

From a broader perspective, at the turning point where the industry moves from technical verification to large-scale delivery and stable work, it is necessary to use ecosystem-based layout to hedge the structural risks of a single-product company and achieve cross-generational development through ecosystem collaboration.