Unitree's Wang Xingxing, Wang Jian, and Zeng Wenjun Discuss Physical AI: GPT-6 Enables Robots to Paint with Zero Training
At the 2026 World Physical AI Conference, Wang Xingxing, founder of Unitree Robotics, demonstrated a test video based on the GPT-6 model: the robot, without complex code and with completely zero training, autonomously dipped paint, washed the brush, and completed a gouache painting. Wang Xingxing judged that physical AI has shown signs of a "breakthrough," in as fast as 1 to 2 years or as slow as 2 to 3 years, and predicted that the "ChatGPT moment of embodied intelligence" will arrive as early as 2027. Wang Jian, director of the Zhejiang Lab, and Zeng Wenjun of Ningbo Eastern Institute of Technology respectively discussed the integration opportunity of AI and robots "merging soul and body into one," as well as the data gap facing physical AI and the foundational role of world models.
Body · AI translation
Article body EM_StockImg_Start
EM_StockImg_End
On October 22, at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Unitree Technology founder Wang Xingxing showed a test video based on the GPT-6 model.
In the video, a reference painting and a blank sheet of paper are laid out in front of the robot. It skillfully picks up a brush, dips it in paint, and, following the reference painting, creates a gouache painting on the blank paper. During the painting process, it even dips water on its own and cleans the brush, switching between different paints, with a very high degree of autonomy.
"We did not write any complex code." Wang Xingxing's tone could hardly conceal his excitement, and he repeated "completely zero training" several times.
As the capabilities of foundation models continue to improve, artificial intelligence is accelerating from the digital world into the physical world. At the conference, Wang Xingxing made a judgment: "Physical AI has already shown signs of being 'broken through.' At the fastest, one to two years; at the slowest, two to three years."
Large models enter the physical world
Wang Jian, director of Zhijiang Laboratory and founder of Alibaba Cloud, believes that the current rapid development of artificial intelligence is largely built on breakthroughs in large language models, especially in text processing capabilities. But human activities such as reading, speaking, and writing are all inseparable from bodily movement and participation in the physical environment, and the complexity that the real world requires models to describe and handle far exceeds that of text.
In his view, artificial intelligence and robotics used to belong to different technological fields, but with the development of computing technology and large models, the two have an opportunity for deep integration. Wang Jian likened this change to "the soul and the body finally finding a chance to become one."
The robot painting video at the conference reflected to some extent the potential for model capabilities to transfer to the robotics field. Wang Xingxing admitted that he had previously underestimated AI's capabilities: "Almost everyone in the world, in terms of how much they value or believe in AI, actually does not value or believe in it enough, including myself." He said that existing model capabilities can, in a certain sense, already solve robots' planning and navigation problems. At the same time, model capabilities are far from their upper limit and will continue to improve rapidly in the coming months or one to two years.
Looking globally, tech giants are racing to lay out physical AI. On October 21, Nvidia announced an expansion of the Isaac robotics development platform, adding tools such as a robot training framework, datasets, and runtime environments, further connecting the process from data generation and model training to validation and deployment. Google DeepMind also released the Gemini Robotics ER 2 model in July this year, improving robots' spatial reasoning, multi-step task planning, and collaboration capabilities.
At present, Unitree Technology is exploring the construction of a self-evolution loop system for physical AI robots, that is, letting AI models write code autonomously, having robots train in simulated environments, then testing them in real environments, and conducting autonomous evaluation and adjustment based on test results, gradually forming a closed loop of autonomous iteration. Wang Xingxing regards this as Unitree's "most important matter this year."
"As the positive cycle is established, the iteration speed of models or robot algorithms will become faster and faster, without the need for (human) participation, and the self-evolution capability will also become stronger and stronger," Wang Xingxing said.
"The ChatGPT moment of embodied intelligence"
may arrive as early as 2027
On the technical level, Zeng Wenjun, vice president of Ningbo Eastern Institute of Technology and a foreign academician of the Canadian Academy of Engineering, said bluntly that the development of physical AI faces a large data gap. Collecting real physical data is time-consuming and labor-intensive, and it also faces challenges such as characterizing physical laws, real-time interaction, safety, and interpretability. Under conditions of limited data and computing resources, whether the technical route of relying on large-scale data and computing power to drive model development is still applicable to physical AI deserves further consideration by the industry.
On this basis, Zeng Wenjun believes that world models are expected to become the foundation for the development of embodied intelligence. Their basic capability is to predict the subsequent state of the physical world based on the current observed state and actions. Therefore, world models can serve as a data engine to generate simulated environments and assist robot training; they can also be used for robot action decision-making, helping robots choose the next action by predicting future states.
Wang Xingxing believes that the core challenge urgently facing physical AI is the alignment between the input and output of AI models and real robots. For example, sensor perception has errors, actuators also produce errors in actual operation, and multimodal large models still have relatively limited ability to correct errors during physical contact. These problems will all affect robots' actual performance in complex environments.
At the same time, the robotics industry needs to make arrangements for large-scale application. "In the future, if many factories, shopping malls, and homes all have robots, (production capacity) must reach at least the tens of millions of units level," Wang Xingxing said. Current industry production capacity still falls far short of the demand for large-scale robot popularization. In addition, the construction of ultra-large-scale server clusters and the manufacturing capacity for computing power chips must also be considered.
Despite the remaining challenges, Wang Xingxing remains optimistic about the timing of breakthroughs in embodied intelligence. At this conference, he once again mentioned "the ChatGPT moment of embodied intelligence," that is, robots being able to successfully complete about 80% of tasks in about 80% of unfamiliar scenarios according to voice or language instructions. He believes that this moment is expected to arrive in 2027 or 2028, and may arrive as early as 2027.
(Source: Shanghai Securities News) Body middle embedded advertisement