_________

Published on Oct. 30, 2026

OpenAI’s GPT-6 Astra can control robots it was never trained for. That surprise has reopened old questions in China. How far can language skills be transferred to actions in the real world? Will large language models and world models eventually merge? Or will they continue to complement each other as separate modules?

Five AI experts from Chinese universities, research institutes and companies discussed these questions at the 2026 World Model Summit Forum in Beijing on October 11, 2026. The Beijing Key Laboratory of General World Models organized the event. The business outlet Mingliang Gongsi published their discussion on October 14 (in Chinese).

Most participants agreed on one point. Language reasoning, visual perception and physical interaction need to grow closer together. The greatest bottleneck today is the shortage of interaction data. Methods such as remotely operating robots can help fill that gap.

On other questions, their views differed. Should integration take place within one model or across a system of several? Do the data available today support a unified architecture at all?

At the heart of the debate are the different capabilities of AI technologies. A large language model, or LLM, such as Astra reads instructions and reasons about a task. A vision-language-action model, or VLA, translates what a robot sees and hears directly into movements. A world model learns how the physical environment behaves and predicts what an action will cause.

Chinese engineers often call the reasoning layer the robot’s “brain” and the movement layer its cerebellum.

“Large language models can be responsible for understanding, planning and coordinating long-horizon tasks,” said Zhang Li, a professor at Fudan University, in Beijing. VLAs execute those intentions. Motion control adapts them to humanoids or dexterous robotic hands. “However, at the current stage, coordination between different components at the system level remains very important.”

The question is whether language is a kind of scaffold on which physical intelligence can be built. Or does language constrain that intelligence? Three camps in China give different answers.

The first camp sees the new Astra model as evidence that language is a scaffold. The highest form of intelligence would be a unified model for all modalities, argued Yu Chao, an assistant professor at Tsinghua University’s Shenzhen International Graduate School. Robots would use smaller models whose capabilities are distilled from the larger ones.

A few industry voices put it more radically. Large investments in VLA and world-model approaches might no longer be worthwhile, said He Yonghao. He is CTO of Xushi Technology and formerly led embodied AI at D-Robotics. Future advances would come from companies developing general foundation models, he argued.

Zhu Zheng, co-founder and chief scientist of robotics startup GigaAI, warned that models like GPT-6 had already absorbed multimodal capabilities and were now encroaching on embodied AI. The developers of frontier models had more talent, computing power and funding at their disposal.

“The wolf is already here. Abandon your illusions,” Zhu said. He sees the next one or two years as a critical window. During that time, embodied AI companies must build a competitive advantage that protects them from attack and that they do not yet possess. Investors call such an advantage a “moat.”

The second camp argues that language constrains robots. “If we want robots to perform real-world tasks, they must have access to sufficiently comprehensive information,” said Jin Xiaojie, a professor at Beijing Jiaotong University’s School of Computer Science, at the Beijing forum.

Operational details and the relationships between objects and their surroundings are difficult to put into words. “Relying exclusively on language models is unlikely to provide the ultimate solution for embodied AI,” Jin said.

Zhang Borui, an assistant researcher at Tsinghua University, supplied the central argument for this position. Language data dwarfs physical interaction data. A model trained on both would therefore mainly learn language. “Under current conditions, allowing different modules to exploit their respective strengths remains a practical choice,” he said.

The third camp in China finds both positions convincing and combines them in a pragmatic plan. The debate reminded Xia Zhongpu, co-founder and CTO of Anyverse Dynamics, of the development of autonomous driving. There, separate systems initially ran in parallel and were integrated later. Language compresses the physical world, while images describe it in detail, he said.

He gave an example. The rule to stop at a red light fits into a short sentence. But in a camera image, the traffic light occupies only a few pixels. Everything else in the image must also be processed. “World models therefore need assistance from language, while language models need world models to provide the missing details,” Xia said.

At the PAIR Conference in Shanghai on September 20 and 21, 2026, a closing panel opened with a provocative question. Do robots still need world models at all? The Shanghai Institute of Physical Intelligence and Robotics organized the conference.

“Not necessarily,” answered Ban Yutong, an assistant professor at Shanghai Jiao Tong University, according to an October 22, 2026 report by the science magazine MIT Technology Review China (in Chinese). A model that predicts realistic images has not necessarily understood the principle of friction, Ban warned.

Perhaps “a complex external model is not necessary,” said Sui Yanan, an associate professor at Tsinghua University. A robot picking up a cup does not need a reconstruction of every object in the room.

Chinese companies, meanwhile, began testing the new Astra model in practice immediately. The robot maker Galbot published the results of an initial test on September 12, just nine days after GPT-6 Astra’s introduction. A team led by Galbot founder and CTO Wang He, who also teaches at Peking University, used Astra to control a simulated dual-arm robot in the RoboDojo benchmark.

The tasks included tidying a table, building towers, or folding clothes, Chinese-language media reported.

Astra completed 13 of 50 trials successfully, a success rate of 26 percent. Astra alone controlled the robot in these trials.

The results were better when Astra was paired with a smaller action model as an “assistant.” Combined with π0.5, pronounced pi-zero-point-five, the open-source action model from specialist Physical Intelligence, the system completed 24 trials, or 48 percent. π0.5 generated the movements. Astra observed, assessed and corrected them. It changed only 14.4 percent of the action-control steps.

Exactly how the two models work together appears to be decisive. An evaluation led by Chinese AI developer SenseNova was published on October 19, 2026. It covered 12 benchmarks and eight real robots. It found that π0.5 alone scored 6.91 points on RoboDojo.

When Astra merely gave it subgoals, the score fell to 5.83. When π0.5 proposed each movement and Astra accepted, edited or rejected it, the score rose to 23.61. With a few examples shown to Astra during the task, the system reached 54.17 points. The brain contributes most when it reviews the cerebellum and least when it only gives orders.

The evaluators also discussed token costs. The SenseNova authors describe the approach as “costly but effective.” On a real robot, they write, “it is implausible to trade longer execution for higher success rates with physical constraints.”

For practical use, they draw a clear conclusion. Astra has “largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.” In a widely shared example, Astra needed more than four hours and US$55 worth of tokens to put a pen into a holder, reported the business newspaper 21 Shiji Jingji Baodao.

A final verdict would be premature. Astra is only a few weeks old, and its successor could move the boundary again. But the experts’ discussions at conferences and the actual robot tests with the new Astra model are currently pointing in the same direction.

Language models expand what robots can understand about their environment. Action models and world models continue to supply what language leaves out. Better integration of the two approaches is emerging as a consensus in Chinese academia and among the country’s physical AI companies.

That distinguishes the Chinese debate from the one in the United States. Sam Altman’s OpenAI expects language models to take over robotics. Yann LeCun’s camp argues that generative language models cannot handle the continuous, noisy signals of the physical world.

Judging by this fall’s discussions, most Chinese experts do not take either side. They assess what each approach does best and ask how the two can be combined at a price a factory is willing to pay.

For China’s robot makers, the question is where they can make the greatest contribution themselves in the future. If powerful language models are available from other companies, their own interaction data, specialized action models and adaptation to particular robot hardware could become more important.

“Today, the essence of embodied intelligence resides in the data. Models are more like distillers and containers,” said Wang Qian, CEO of X Square Robot, at the INCLUSION Conference on the Bund. The hope is to sustain a competitive advantage over time based partly on the sheer size of China’s manufacturing industry.

In China, the thinking is that GPT-6 Astra has made the value of language as an intermediary for planning long-horizon tasks clearer. Language remains especially useful for high-level relationships and reasoning. Visual information, however, is needed for concrete details. Until a single model exists that perfectly unites both capabilities, the brain and the cerebellum must learn to work together.

For now, there is no clear winner in the Chinese debate over language models and world models. The focus is on how their respective strengths can be combined.