OpenAI confirms Liang Wenfeng’s “prophecy”

Letter AI

2026.11.20 08:05

· From Beijing

Full text 4468 words

Text | Letter AI

GPT-6.2 Astra topped the real robot arm test, taking the capabilities of general large-model direct-drive robots a step further.

According to the AA-Robotics v0.1 list released by the independent evaluation agency Artificial Analysis, GPT-6.2 Astra scored 79 and completed 6 out of 8 physics tasks.

Not only OpenAI’s flagship model, but also the relatively weak GPT-6.2 Sol also received a high score of 78 points.

In this benchmark test, the model does not allow special training, and zero-sample settings are used throughout. The model directly controls a 6-axis robotic arm to move real objects by writing code.

In the past, robot manufacturers generally believed that vertical robot models had advantages in robot adaptation and force control, and were more suitable for robots than universal models.

It now seems that with the development of general models, after performance iterations, general models may replace the status of vertical models.

Liang Wenfeng has expressed a similar view. He believes that breakthroughs in embodied intelligence can be based on the continuous evolution of general model capabilities.

In Liang Wenfeng's AI route, the evolution of the capabilities of the general model itself is the main line leading to AGI, and embodied intelligence is the extension of this route.

01

On November 13, foreign media reported that OpenAI disbanded its independent robotics team and merged the original robotics team into the company's main research team. The application robotics and hardware teams were adjusted to other departments.

OpenAI confirmed to the media that the reason for the adjustment is that the general model performs well in controlling robots.

Researchers found that after connecting GPT-6 Astra to the robot, it was able to participate in the execution of some specific tasks and performed surprisingly well.

In fact, after the release of GPT-6 Astra, there has been a round of attempts to use Astra to drive robots. For example, real-machine testing by the independent evaluation agency Robocurve shows that Astra achieved a 95% success rate when using a two-arm robotic arm to perform the task of "grabbing blocks and placing them in a bowl."

Such experiments give universal model-controlled robots more realistic basis. In this context, OpenAI merged the independent robotics team established less than half a year ago into the main research team, and the researchers turned to the development of general artificial intelligence models.

This means that OpenAI has changed its technical route. Now that the general model has begun to control the robot, there is no need to develop a separate special "spatial intelligence" model for the robot.

In fact, this is not the first time OpenAI has adjusted its direction on the robotics technology route.

As early as 2019, OpenAI established a robotics team, trying to use reinforcement learning and other methods to let robots learn tasks such as grabbing and placing objects. But by 2020, the team was disbanded.

Its co-founder Wojciech Zaremba once explained that robot data is limited and AGI can be implemented more efficiently through other methods. Later, OpenAI focused on large language models and launched ChatGPT in 2022.

With the development of multi-modal models, OpenAI has turned its attention to robots again.

In 2024, OpenAI will cooperate with the humanoid robot company Figure AI to try to apply large model capabilities to robots.

At that time, Peter Welinder, vice president of OpenAI, said that the cooperation with Figure was to explore the level of intelligence that humanoid robots can achieve with the support of large-scale multi-modal AI models.

Later, the Figure 01 demonstration video released by Figure AI showed that the robot could complete simple item delivery according to human instructions. In this collaboration, OpenAI’s model provides visual understanding and language interaction capabilities.

However, a year later, Figure AI announced the end of the cooperation and turned to independent research and development of robot AI.

And OpenAI’s robot has an immortal heart. In early 2026, OpenAI researchers began to explore another technical route: starting with video generation models like Sora.

As a video generative model, Sora is able to understand the relationship between objects, space and motion, giving it a certain understanding of how the physical world works.

OpenAI’s turn is not an isolated case. At the beginning of 2025, the mainstream route for embodied intelligence was still the VLA model. By the end of the year, as VLA's shortcomings in cross-scenario generalization and other aspects emerged, the industry began to turn its attention to the world model.

In June of this year, two months after Sora was shut down, Sam Altman announced the establishment of the robotics department "OpenAI Robotics" and shifted the Sora team as a whole to robot research and development. Led by Aditya Ramesh, the father of Sora and founder of DALL-E, the world simulation technology accumulated by Sora has been transferred to robot training. The short-term focus is on industrial auxiliary robots, and the long-term goal is consumer-grade humanoid robots.

But the route soon changed again.

Just three months after the establishment of this department, in September this year, OpenAI released GPT-6 Astra. OpenAI found that the general model itself is already capable of participating in robotic tasks. Compared with world simulation technology that specializes in robots, there is a way to directly use the existing visual understanding, reasoning capabilities and task planning capabilities of general models.

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, commented that this means that OpenAI believes that robotics "is no longer a special problem."

After a detour on the road of robotics, OpenAI turned its attention to general models.

02

OpenAI's changes are not an isolated case. The performance of universal models on robotic tasks is causing the industry to review this technology route.

The Astranomy research released by NVIDIA in October directly uses GPT-6 Astra to control simulated robots. In the RoboLab-120 test, the direct control scheme achieved a mission success rate of 66.3%.

Deepu Talla, head of NVIDIA's robotics department, said in a recent interview that the emergence of GPT-6 Astra is a major breakthrough in promoting multi-tasking capabilities for robots.

Claude Fable 5.2 released by Anthropic also demonstrates similar capabilities. The model-controlled robot completed a variety of tasks, such as solving Rubik's cubes, unscrewing bottle caps, and drawing robots. In the RoboDojo Sim simulation benchmark, the average success rate of Fable 5.2 reached 64.9%, and the score of GPT-6 Astra under the same test framework was 29.8%.

Although this technical route is not yet mature, it at least shows that the universal model is becoming a route worth trying in embodied intelligence research.

In China, robotics companies are also paying attention to this change.

Yushu is exploring the use of large models in robots. Wang Xingxing even bluntly said: "When GPT-7 comes out next year, it may directly solve the problem of embodied intelligence."

Previously, Yushu was both promoting the VLA model and exploring the world model. Wang Xingxing once mentioned that last year Yushu picked up the world model based on video generation again. Yushu has released the universal VLA model UnifoLM-VLA-0 and the WMA (World Model-Action) model.

But recently, Yushu has begun trying to use GPT-6 Astra for robotic experiments.

Wang Xingxing once mentioned that without writing complicated programs in advance, the robot can complete the painting only through prompt words; for example, the robot not only recognizes the chessboard, but also accurately places the chess pieces in the corresponding position. Wang Xingxing emphasized that these experiments were completed under "zero training" conditions.

Zhiyuan Robot is also exploring a similar direction.

Recently, Zhihui Jun's team released a paper. The core of this paper is to connect a large model that can chat with a robot body that can perform actions. The model can not only understand verbal instructions, but also use the robot's body to complete tasks such as pointing to real objects, body movements, and facial expressions.

This means that robotics companies are trying to "connect" the capabilities of general-purpose large models into robot bodies to promote embodied intelligence.

This also gave Liang Wenfeng a new reference for his previous judgment.

When world models became a popular direction in the field of embodied intelligence, Liang Wenfeng appeared relatively restrained on this technical route. He values ​​​​the evolution of the capabilities of the general model itself more and regards it as the main line leading to AGI.

In Liang Wenfeng’s AI route concept, after Agent, we must first solve the continuous learning ability of the AI ​​​​model, and then reach the singularity that can iterate by itself, and then embodied intelligence.

In other words, he hopes to make the general model itself strong enough first, and then extend this ability to the physical world.

This bears some resemblance to OpenAI’s latest tweaks. Both point to the conclusion that embodied intelligence does not necessarily require training an independent set of robotic intelligence from scratch. The evolution of general model capabilities may also become a path for robots to move towards embodied intelligence.

And neither company is stopping at the model level.

While OpenAI is adjusting its robotics business, it has invested in a number of robotics companies. DeepSeek has also begun to extend into the field of embodied intelligence, participating in the placement of Yushu Technology, reaching a three-year strategic cooperation with Yushu, and investing in companies such as Winverse AI and Lige Wanwu.

03

The integration of general large models into robots has become a new trend, but this does not mean that VLA models and world models will disappear.

In the past few years, the mainstream technology route of embodied intelligence has undergone several changes, from VLA model, to world model, to general large model. But no one ever really replaces anyone.

The first one to receive attention was the VLA model. The so-called VLA (Visual-Language-Action Model), simply put, allows the model to process camera images and human instructions at the same time, and then convert the understanding results into robot actions.

The emergence of this route is related to the development stage of large models at that time. At that time, the large model had certain language and visual understanding capabilities, but it could not directly control the robot. The robot's own data is limited, and it is difficult to train a robot model from scratch.

In July 2023, Google DeepMind released the VLA model RT-2, which is able to learn from network and robot data and convert this knowledge into universal instructions for robot control.

The VLA route provided a feasible solution at the time. However, VLA lacks generalization ability and autonomy. As a result, the robot cannot adapt to complex real-world environments.

In the real physical world, robots need to deal with continuous physical changes. The robot not only needs to know what it should do now, but also determine how the world will change after doing so.

Therefore, when the video generation model came out, manufacturers generally believed that the video generation model was expected to replace the VLA model.

In February 2024, OpenAI released Sora and proposed in a technical report that large-scale video training has become the path to building a physical world simulator.

There are tons of videos of human activity on the Internet. How humans pick up things, move things, walk, organize things, etc. are all recorded in videos. If the model can learn object motion, spatial relationships, and action consequences from these videos, there is an opportunity to transfer this ability to robots.

But video generation is not the same as being able to control a robot. As a result, the industry began to try to combine world prediction and action generation to form a world action model (WAM).

In other words, VLA focuses on mapping from observations to actions, while the world action model attempts to incorporate predictions of the future into action decisions.

However, the world model still faces many problems, and the simulated physical processes may not be accurately reproduced in the real environment.

After Astra, the industry began to explore another possibility. Can general-purpose large models directly undertake the task understanding, planning and decision-making of robots?

HomeBody, the embodied intelligence research project released by the Stanford team in conjunction with the California Institute of Technology, provides an example. This study attempts to answer a question: If the universal large model is smart enough, does the robot also need a VLA model as an intermediate layer between the large model and the body?

The answer given by HomeBody is: Within a certain range of tasks, it is not necessary.

After the researchers connected GPT-6 Astra to the Yushu G1 humanoid robot, the robot directly called the pre-built robot skill library. After the robot entered a kitchen it had never seen before, it completed sorting and searching for items according to human instructions.

The same is true for Fable 5.2, which was later released by Anthropic. Fable 5.2 has significantly improved its ability to control robots, even for those robot tasks for which it has never been specially trained.

This means that as the model’s reasoning capabilities increase, some robot skills that originally required specialized training may begin to emerge from the general model.

A core issue of embodied intelligence is to allow robots to understand the environment and determine the possible consequences of actions. The idea of ​​the world model is to learn from videos and then simulate the future. The idea of ​​a general large model is to use existing reasoning capabilities to analyze tasks and then use them for task planning.

The technical ideas of the two routes are different, but they point to the same goal: to let AI take action in the real world.

Respect for the original work, if you are qualified, you should be appreciated.

According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.

Express

08:05

Nvidia closes in on acquiring diffusion AI startup Inception for more than $500 million

08:04

EU approves Hungary's 5.4 billion euro defense loan plan

08:01

Wang Yi introduces the situation of the meeting between the heads of state of China, Russia and the United States

07:59

MSCI will add Anthropic to its large-cap index on December 1

07:39

Anthropic CEO: The company hopes to obtain an investment-grade credit rating in the future

07:36

Anthropic CEO responds to US$900 billion AI investment commitment: It is not large compared to existing cash reserves and revenue

07:34

SpaceX reportedly plans to issue at least $32 billion in AI chip-related debt

07:32

Anthropic rings the Nasdaq opening bell, setting a record for the largest IPO in history

07:20

U.S. stock market closing comments: The three major stock indexes all fell more than 2% after opening high and moving low. Anthropic closed up 17% in its debut.

07:14

Trump says he won't 'rush' into action against Iran before New Year

07:08

The net value of the fund reveals the direction of position adjustment, and institutions explore the certainty of the next stage.

2026-11-19 23:11

Director Zhu Lidan sues Wahaha Commercial Co., Ltd.

2026-11-19 23:02

Domestic commodity futures markets mostly closed higher at night, with black products leading the gains.

2026-11-19 23:01

International Energy Agency: Middle East conflict may reshape global energy landscape

2026-11-19 22:59

SpaceX plans to issue at least $32 billion in AI chip debt on Thursday

2026-11-19 22:55

Anthropic is expected to open at $99.60 on its first day of trading

2026-11-19 22:52

Livermore’s leading index of Chinese concept stocks expanded to 2%, while Pinduoduo rose nearly 13%.

2026-11-19 22:51

About 15% of shares in Anthropic IPO sold to retail investors

2026-11-19 22:50

亚太经合组织互联互通蓝图

2026-11-19 22:46

Anthropic的IPO未向多数机构买家配售股份

扫描下载App