Economic Observer reporter Zheng Chenye

We didn't specifically write programs for these actions. It saw a paint tray next to it and went over to wash the pen clean before continuing to draw. "On October 22, Wang Xingxing, founder and chairman of Yushu Technology (688836. SH), introduced an internal test at the 2026 World Physical AI Conference, allowing OpenAI's GPT-6 to directly control the robot to draw without specialized training.

Wang Xingxing sighed that everyone's faith in AI is actually not enough, including himself. On November 4, at the 2026 World Internet Conference Wuzhen Summit, he also said that the world's leading multi-modal large model has solved the problem of task planning very well, and next year is probably the critical point for the robot industry and robot models.

The multimodal big model mentioned by Wang Xingxing mainly refers to general big models such as GPT, Gemini, Claude, etc. This kind of model was originally a big language model. It was originally used to predict the next word based on the above, and then gradually added image and video processing capabilities. The training data mainly came from the Internet.

OpenAI's GPT-6 Astra, released in September this year, extends the capabilities of such models from task planning to direct control of robots.

In a paper released on September 21st, the RoboDojo team, a robot evaluation platform, and researchers from institutions such as the University of Hong Kong and Tsinghua University pointed out that without fine-tuning for specific tasks, Astra's average success rate in RoboDojo simulation tasks was 22.48%, while the previous generation GPT-5.5 was only 0.88%. This achievement was once higher than all publicly available specialized robot models on the RoboDojo list.

The emergence of high-performance general-purpose models such as GPT-6 Astra has raised some doubts about whether the World Model route, which specifically allows AI to understand the physical world, has been falsified. Or, to put it another way, since general large-scale models can also control robots, with the continued enhancement of model capabilities, it seems unnecessary to specialize in world models.

Is this really the case?

What can a universal big model replace

The world model originally referred to a model that can predict the next state of the world based on the current state and a given action, such as predicting whether a robotic arm will push a water cup and whether it will fall onto the table. In the past two years, the world model has become popular, and its meaning has been constantly changing with the advancement of AI.

As early as 1943, British psychologist Kenneth Crick proposed that there are small models of the real world in the human brain that can be used to predict outcomes before taking action.

In the late 1980s and early 1990s, this idea entered the field of neural network research. Researchers David Ha and J ü rgen Schmidhu ber published a paper titled "World Mod els" in 2018, which allowed AI to first learn the changing patterns of the environment and then train the learned models. As a result, the name "world model" gradually became popular in the field of deep learning.

In the following years, the world model was mainly the research direction of the academic community. Turing Award winner Yang Likun published a paper in 2022 advocating the construction of machine intelligence with world models as the core. After big language models became mainstream in the industry, he still publicly criticized such models for lacking understanding of the physical world.

Starting from 2024, the popularity of world models will significantly increase, and video generation, robotics, and autonomous driving all require such capabilities.

In February 2024, OpenAI released the video generation model Sora, with a technical report titled 'Video Generation Model as a World Simulator'. In September of that year, World Labs, founded by Chinese scientist Li Feifei in the field of AI, made its public debut and raised $230 million. Its goal was also to enable AI to understand the three-dimensional physical world.

In January 2025, Nvidia released the world model platform Cosmos, which generates training data for robots and autonomous driving. NVIDIA founder and CEO Huang Renxun said that the "ChatGPT moment" in the robotics field is coming soon.

The world model track was immediately sought after by capital. After leaving Meta at the end of 2025, Yang Likun founded AMI Labs, a world modeling company, and announced in March this year that it had raised $1.03 billion. The company's proposition is that true intelligence does not begin with language, but with the world. In September, AMD announced its intention to acquire World Labs for approximately $8.2 billion.

According to statistics from market research firm Venture Capital, there will be 23 financing events in the fields of physical AI and world modeling in 2025, and over 100 events have occurred since 2026.

In an article published by World Labs in June this year, it was stated that the world model is one of the most ambiguous terms in the current AI field, and it is divided into three categories based on output results: first, generating images for humans to see, second, simulating physical states, and third, determining the next action.

For these three types of work, the general big model has now been involved.

In terms of generating images, the video generation model launched by General Motors has already possessed this capability. However, Lin Meibing, Chief Analyst of ICTIME, told Economic Observer that realistic graphics do not necessarily mean that the model understands physical laws. Video generation models that do not accept action input cannot strictly be considered world models.

In simulating physical states, the approach of general large models is to call on existing tools. In a research report released by Zheshang Securities on October 21, 2026, it was pointed out that Astra has been able to use the 3D software Blender to restore real space, and Google's Gemini 4 Argon can adjust parameters such as friction, and then use the physics engine to calculate contact and motion. That is to say, a general large-scale model does not necessarily require mastery of physical laws.

In terms of deciding the next action, the ability of the general large model has undergone the greatest changes recently. According to simulation tests released by NVIDIA's Seattle Robotics Lab in October, the success rate of Astra, which did not undergo robot specific fine-tuning, was 66.3% among 120 grasping and placement tasks. NVIDIA's robot model developed based on the world model Cosmos was 36.8%, and the π 0.5 model specially trained for robots by American company Physical Intelligence was 28%.

But in terms of execution speed, the order of the three models is exactly opposite: in the above test, when Astra directly outputs actions, a single call takes an average of 18.56 seconds; This model from Nvidia takes about 0.71 seconds; π 0.5 only takes about 47 milliseconds.

Robot control requires high speed. Yao Maoqing, co president of Zhiyuan Robotics and president of the embodied business system, previously told Economic Observer that Zhiyuan's control system is divided into three layers, with the lowest layer communicating with the motor at a rate of 1000 times per second; The middle layer is responsible for motion planning, 10 to 20 times per second; The top layer is responsible for task scheduling. That is to say, for robots, the speed of the universal large model cannot keep up with the lower two layers, and it is currently more suitable to be placed on the top layer.

A product manager from a leading robotics company in Shenzhen told Economic Observer that in terms of specific methods, the company allows the big model to understand tasks and disassembly steps, but still leaves the specific actions to their trained action models for processing.

In addition to speed, operational accuracy is also a weakness of general large models. On November 5th, Wang Xingxing stated during the 9th CIIE that the biggest challenge at present is not the insufficient understanding ability of the model, but rather the "significant deviation in the last few millimeters" when executed in the physical world.

In September of this year, the independent evaluation team Robocurve announced a real machine test, in which Astra put blocks into a bowl and succeeded 19 times out of 20, but only succeeded 2 times in inserting precision puzzles 20 times.

The shortcomings in accuracy are related to the training data. Video learning can let the general large model know that a person has picked up a cup, but it won't tell it how to move the robotic arm, "said the product manager mentioned earlier. The movement of the robotic arm relies on real machine data, which is the data collected during the actual operation of the robot.

Huang Xiaohuang, chairman of 3D design software company Qunhe Technology (00068. HK), said at the Wuzhen Summit on November 3 that the reason why the big language model is the strongest is that the Internet has left a huge amount of text data that can be captured for free, but the data in the physical world can only be collected by people or robots, and the cost is still high.

Another weakness of the universal big model is predicting the consequences of actions in practical operations. In a paper released on October 20th, researchers from institutions such as Nanjing University, Nanyang Technological University, and Peking University pointed out that there is a significant gap between Astra's understanding of the task and its actual execution, with problems focusing on predicting action consequences, accuracy, and coordinating complex actions.

And predicting the consequences is the main task of the world model, and General Motors is also developing the world model itself.

On October 27th, Google DeepMind released a research preview version of the world model Genie4. The general large model Gemini is responsible for disassembling targets and adjusting plans, while Genie4 is responsible for simulating what will happen next. The two work together to simulate a humanoid robot playing checkers.

At least for now, predicting the consequences still requires the world model to complete.

Pure world model company faces increasing pressure

The ability to predict consequences has its value, but there seem to be not many buyers who specifically pay for the world model.

The trial and error in the physical world is different from that in the digital world. If a large model writes the wrong code, it can be modified and run again; If the robotic arm knocks over the water cup, it cannot be retracted due to an error. The higher the cost of trial and error, the greater the value of predicting in advance.

However, this ability may not necessarily need to be provided by specialized world model companies: robotics companies and car manufacturers often develop it themselves or use open models; Chip companies open up their models to drive computing power demand.

A common practice in the robotics industry is to incorporate world models into their own trained action models. Guo Yandong, founder and CEO of the robotics company Zhifang, once gave an example that the steps of brewing tea rely on language models, and the water cup may fall off near the table. Such predictions rely on world models, and if the world models are not merged, many tasks cannot be completed at all.

At present, robot companies mainly use world models for two things: one is to deduce possible impacts before executing actions, and the other is to generate simulation data. In both cases, the world model is only a part of the entire robot company's model.

The aforementioned product manager stated that there are already free world models on the market, and the company is more willing to invest funds in fine-tuning data and models rather than specifically purchasing world models.

One of the attractions of the world model is the use of low-cost video data, which reduces the amount of real machine data collected, but after comprehensive calculation, the cost may not necessarily be lower.

According to the product manager, one way is to train the world model with Internet video, which requires millions of hours of data, and then fine tune with tens to hundreds of hours of high-quality robot data; Another way is to directly train the action model, which requires thousands of hours of data, most of which are real machine data, with an hourly cost of about 300 to 1000 yuan. The former has a lower unit price and larger usage, while the latter has a higher unit price and smaller usage. The total cost of the two methods may be similar.

Robot companies pay more attention to real machine data than video data. Mo Lei, Vice President of Zhifang, recently stated in an interview with Economic Observer that real machine data plays a more crucial role in improving the subsequent performance of the model, and "data diversity is more important than data quantity".

The situation of car companies is similar, with leading companies in the field of intelligent vehicles mostly conducting independent research or developing based on open models.

The world model X-World released by Xiaopeng Motors (09868. HK) in April this year is based on Alibaba's (09988. HK) open-source video generation model WAN2.2; The new version of intelligent driving launched in September has enabled the world model X-Foresight to be loaded for the first time, and it is possible to predict what will happen in the next 6 seconds.

Baidu Group (09888. HK) released the world model Apollo World on November 3 for the test and verification of Apollo Go.

Wu Peidong, Managing Director of Wanchuang Investment Bank, told Economic Observer reporters that open models and tools shorten the start-up time, and top car companies will still independently develop core technologies, with second - and third tier enterprises benefiting more significantly. In his opinion, the weight of intelligent driving in car purchasing decisions has increased, but as an independent payment option, consumers' willingness is still weak. They are more willing to pay for using it with peace of mind and for someone to take responsibility after an accident, and are less willing to pay for "thinking". This also leads to a lack of motivation for car companies to pay for external models.

Outside of robotics and automotive companies, chip companies focus on world models primarily for computing power needs.

Nvidia has opened up the world model Cosmos to the industry, and its staff told Economic Observer reporters that opening up these models can help the industry achieve a "60 point" score for basic models. Car companies can then use a small amount of computing power and their own road test data for fine-tuning, but the use of these open models across the industry will ultimately translate into a sustained demand for cloud training and simulation computing power.

AMD has chosen to directly acquire WorldLabs, and the transaction is expected to be completed by the end of the year. AMD Chairman and CEO Su Zifeng stated at the earnings conference on November 3rd that this acquisition will allow AMD to gain a deeper understanding of the evolution of cutting-edge models. In Lin Meibing's opinion, AMD's focus is mainly on the model team. After integrating the team into the system, chip companies can grasp the requirements of the next generation model for video memory, bandwidth, and latency earlier.

Robot companies and car companies rarely pay specifically for external world models, and chip companies open up their own world models, which puts commercialization pressure on world model companies in the market. In addition, the investment in training world models is also considerable.

The domestic world model company has excellent vision. Its co-founder and chief scientist Zhu Zheng publicly stated in September that if the training data needs to reach millions or even billions of hours, "it may take hundreds of billions of yuan to train (train) a truly universal world model". Lin Meibing also believes that world models consume a lot of computing power, and the computing power scale of startups is difficult to compare with that of general large model companies.

In fact, General Motors is also accelerating its entry into this field. OpenAI announced in March this year that it would cease Sora's consumer applications and API services, and Sora's research team will shift to world simulation to promote the development of robotics.

OpenAI CEO Altman recently publicly stated that OpenAI will manufacture its own humanoid robots and is expected to produce a demonstration prototype by 2027. During the interview, the reporter also learned that self-developed hardware can help large model companies obtain high-quality training data more efficiently.

In this situation, Zhu Zheng believes that there is not much threshold for General Motors to shift to world models, and the time window left for pure world model companies is not very long. It must fulfill the expectations of the public within one to two years. Looking ahead to the next five years, my biggest risk or concern for the world model is whether the commercial closed loop can work, "said Zhu Zheng.

(Author Zheng Chenye)