Wang Xingxing Discusses "Trust AI as Much as Possible" Twice in Ten Days, Unitree Accelerates Building the Embodied "Brain"

In the eyes of most people, Unitree Technology has always had advantages in hardware and motion control within the industry, while its embodied intelligence large models were relatively weak in the past. In the past two years, the company has also been actively laying out in this area. Especially this year, Wang Xingxing has mentioned "embracing AI" multiple times in public statements, and recently even admitted that "the biggest mistake I've made over the past decade or so is not believing in AI enough."

On October 22, at the 2026 World Physical AI Conference and Unitree Partner Conference, Wang Xingxing gave a speech of about 30 minutes on the theme of "Believe, believe in AI as much as possible." He revealed that within the company, he has reminded employees to value AI for at least four or five years, and has paid attention to the stocks of several leading AI companies.

"If everyone had believed in this back then, the stocks they invested in might have risen tenfold or even dozens of times. But in reality, many people didn't think of this." Wang Xingxing mentioned that very few people have truly seen the effects of the most advanced AI. Even within leading companies like OpenAI, not necessarily many people truly understand AI's potential and the extent to which it will develop in the coming years.

Wang Xingxing also shared the capabilities AI has already demonstrated, such as tackling frontier mathematical research problems, software reverse engineering and vulnerability discovery and exploitation, high-fidelity video generation, and robotaxis. In his view, AI will upend the future of everyone's life.

At present, these capabilities mostly solve problems in the digital world.

At the conference, what Wang Xingxing wanted to talk about more was whether AI models can solve problems in the physical world. Judging from the content displayed on site, similar to the focus of industry observers, he is also testing exactly which tasks AI models can enable robots to complete and what the success rates are.

Wang Xingxing introduced some progress in current internal testing. "For example, letting GPT-6 directly control a robot to paint with completely zero training. After the model is deployed to the robot, it can try to directly control the robot to complete tasks." He said that something interesting happened during the test: the robot was able to change colors on its own and also clean the brush. "We didn't write programs specifically for these actions. When it saw a paint palette for washing brushes nearby, it went over and cleaned the brush itself, then continued painting."

He believes that in a sense, some of GPT-6's capabilities can already solve part of the robot's planning problems.

"In the coming years, large model companies will pay increasing attention to machine intelligence and the robotics industry. At the fastest, maybe one to two years; at the slowest, maybe two to three years, everyone will see more progress." Wang Xingxing revealed that they are also developing their own models and cooperating with large model companies at home and abroad to jointly push physical AI faster across the critical point.

As for this moment, Wang Xingxing has mentioned the ChatGPT moment toward embodied intelligence, that is, in 80% of unfamiliar scenarios, through voice or text instructions, robots can successfully complete about 80% of tasks.

Facing the rapid iteration of AI large models, as a technology company, Wang Xingxing believes that consideration should be given to directions such as large-scale manufacturing capacity for robot hardware, advancing the implementation of physical AI models, ultra-large-scale computing power servers, and the manufacturing of computing power and storage chips.

According to his introduction, Unitree Technology has already been conducting internal preliminary research to advance the self-evolution of physical AI robots, that is, to establish a complete loop. They will let AI models collect papers on their own and write code on their own, first training in a simulation environment and then testing in a real environment; the test results will be scored by the AI model, and people can also participate in scoring, and then the scoring results will be fed back to the model so that it can rewrite code and adjust plans. As a positive loop takes shape, the iteration speed of models and robot algorithms will become faster and faster.

"I'm really afraid of wasting every day. Now there are many things I want to do every day. Whether in the robotics field or other related industries, there are many things worth doing, and there are also many new things to learn." At the end of the speech, Wang Xingxing also said that this industry is very hot, and the future market space is large and the prospects are very broad, but competition is also extremely fierce.