OpenAI Disbands Robotics Team Again as Divergence Emerges Between General-Purpose Large Models and Specialized VLA Approaches
OpenAI has acknowledged that it is once again disbanding its independent robotics research team, breaking up and reorganizing several dozen people: core algorithm researchers are being merged into the main model research group led by Mark Chen, while the applications and hardware teams are being placed under Greg Brockman's Applied division. The report states that OpenAI's rationale is that general-purpose frontier models such as GPT-6 Astra have made leapfrog capability gains, making an isolated specialized robotics foundation model no longer necessary; Sam Altman is still preparing a self-developed humanoid robot body to debut in 2027. The article also compares tests such as Claude Fable 5.2 achieving a 64.9% task success rate in RoboDojo Sim, and outlines the divergence among the three approaches: general-purpose models, VLA, and world models.
Why read this
AI score 62
The article reviews the organizational restructuring behind OpenAI's disbanding of its robotics team and compares the divergence between general-purpose large models and specialized VLA approaches in embodied intelligence.
Topics
Body · AI translation
OpenAI's robotics team disbanded! Six years ago it lacked data; today it blames models evolving too fast. 37°21 1 hour ago. Whether the robot brain "descends from the sky" or "grows from the ground" is a question.
Last week, at the Conference on Robot Learning (CoRL) held in Austin, USA, the booth No. 208 reserved by OpenAI was empty, without even the company's logo.
Subsequently, Fortune magazine reported that OpenAI had once again disbanded its independent robotics research team, and OpenAI confirmed this. Saying "once again" is because back in 2020, OpenAI had personally shut down a robotics lab once, with the reason given at the time:
Lacking sufficient physical interaction data, robot R&D progress was too slow, and the company needed to concentrate its efforts on large language models.
Six years later, history seems to repeat itself, but the situation has completely reversed.
This disbandment is not an exit; in fact, OpenAI intends to break up and reorganize this team of dozens of people: core algorithm researchers will be directly merged into the main model research group headed by Chief Research Officer Mark Chen, while the applications and hardware teams will be placed under the Applied department led by President Greg Brockman.
For this round of organizational adjustment, OpenAI's core logic is the leap in capability of the new generation of general frontier models (GPT-6 Astra): general large models are already powerful enough to directly drive physical hardware across domains, and the industry no longer needs an isolated, dedicated robot foundation model.
Unitree's test of connecting to GPT-6 Astra, image source: Unitree Robotics
Even so, Sam Altman is still busily preparing a self-developed humanoid robot body, with the goal of a public debut in 2027.
But is this another huge bubble of general large models blowing into the physical world, or has embodied intelligence truly reached an evolutionary watershed? Coincidence OR inevitability? The "physical intelligence" spillover of general large models
Regardless of whether this is another conceptual narrative from OpenAI, general large models are indeed influencing the evolutionary path of embodied intelligence. As early as July, Anthropic (hereinafter referred to as "A Company") found that the new generation of general models could already complete some tasks that previously required specially developed robot policies, through visual observation, writing control programs, and calling motion control tools.
The emergence of GPT-6 Astra has set off heated discussion and exploration across the entire robotics industry. Experiments by related companies such as Nvidia, Unitree, and Galaxy General also intuitively show the improvements that general large models bring to embodied intelligence—though of course there are limitations.
Even more critically, A Company recently dropped another bombshell. In early November, A Company released its new-generation flagship model Claude Fable 5.2, and one result even left many robotics scholars incredulous:
In RoboDojo Sim, a recognized simulation benchmark for evaluating dual-arm robots completing complex physical operations based on pure visual control, Claude Fable 5.2 achieved a 64.9% task success rate.
What does this mean?
The previous generation Fable 5.1 scored only 27.6%, and Google Gemini 4 Argon, previously widely regarded by the industry as excelling in multimodality and motion planning, scored 34.2%. OpenAI's highly anticipated GPT-6 Astra scored only 29.8% in this test.
Fable 5.2 not only left its peers more than two times behind, but more importantly, it even outperformed the vast majority of vertical AI systems specially trained for robot manipulation.
You should know that Fable 5.2 was not developed for robots at all. It has not undergone millions of hours of robotic arm teaching trajectories; it is essentially still a "digital brain" soaked in massive amounts of text, code, and multimodal image/video data.
Image source: X
The results of blind tests on real machines were even more unexpected. Without any fine-tuning, Fable 5.2 took over the robotic arm relying on zero-shot generalization, not only completing fine movements such as unscrewing bottle caps and solving a Rubik's cube, but even, in a third-party test, controlling the robotic arm on site to complete a full game of 62-move checkers.
Even when a grasped piece slipped and collided due to mechanical tolerance, Fable 5.2 could autonomously perform reverse compensation and correction actions based on tiny changes in the visual image.
Before this, most robot manufacturers believed that the physical world has extremely strict physical laws, friction, inertia, and contact stiffness, and that it must rely on dedicated vision-language-action (VLA) models, high-frequency simulation reinforcement learning, or world models.
But the reality is that general large models, relying solely on scaling and multimodal pretraining, seem to be naturally emerging with an implicit understanding of causal relationships in three-dimensional space.
This also explains why OpenAI dares to withdraw its independent robotics research group. When the evolution speed of the brain far exceeds expectations, continuing to close the door and tailor small models for robots begins to seem inappropriate in both commercial and technical efficiency. Embodied intelligence has come to this day—where exactly are the divergences in the routes?
To understand exactly what general large models have brought to robots, we must first review the changes in the technical paths of embodied intelligence over the past few years.
At first, robots still ran traditional C++ control algorithms and trajectory planning, with large models in the cloud, breaking down human natural language instructions into a series of preset function calls.
Then came the end-to-end VLA wave. The industry tried to fuse vision, language, and action into one large model, taking camera images as input and outputting joint angles or end-effector trajectories that map directly to robotic arm motors. For the first time, robots learned to look at things and act directly, like humans, marking a big step forward in general-purpose generalization. At the same time, the combination of System 1 and System 2 into a "fast-slow dual system" also became necessary.
But this year, the World Model has attracted increasing industry attention.
Compared with directly predicting actions, world models place more emphasis on understanding and predicting changes in the physical world. For example, when a robot is about to push open a door, the model needs to predict what will happen to the door and the surrounding environment after forces of different directions and magnitudes are applied. This capability can be used both for action planning and to help robots learn in simulated environments, reducing reliance on expensive real-world data.
At the same time, VLA and world models are not opposed to each other. NVIDIA's GR00T 2, announced in November, adopted the World Action Model (WAM) architecture, predicting not only robot actions but also the visual changes after the actions occur. Wang He, founder of Galaxy General, likewise advocates fusing VLA and world models to learn richer physical interaction capabilities.
Image source: X
But the progress of general large models has, after all, given a new answer.
In the past, it was widely believed that System 2 required specialized training in robots' semantic understanding, but both Claude Fable 5.2 and GPT-6 Astra are proving that general large models may be the best System 2. Not only do they know encyclopedic common sense about all things, and possess astonishing long-horizon planning capabilities, but as their multimodal capabilities evolve, with just a few glances at camera images on the other side of the screen, they can work out most of the relative distances, occlusion relationships, and operation sequences in three-dimensional space.
Just one day after Fable 5.2 was released, Google also launched its new-generation dedicated VLA model Gemini Robotics 2.5, while using the general model Gemini 4 Argon to generate simulated environments and analyze failure causes to accelerate robot model training.
This perhaps better reflects the true state of the industry today. General models, VLA, and world models still each have their own areas of strength, but the boundaries between them are becoming blurred.
Which one will ultimately be more effective still requires real-world validation. Is the brain "dropped from the sky" or "grown from the ground"?
In fact, regarding the true value and positioning of general large models in embodied intelligence, the entire industry is continuously engaged in a clash of ideas. Different perceptions not only concern technological beliefs but also directly determine where each company's real money is spent.
In February this year, Dario, founder of Company A, said in an interview with The New York Times that although manipulating real robots is more complex than playing video games, from the perspective of the intelligence problem itself, there is no fundamental difference between the two.
Sam Altman and Figure AI founder Brett Adcock are both representatives of the "general-purpose supremacy" camp. In their view, intelligence is the only essential problem, and hardware is nothing more than standard hard labor in industrial manufacturing. With 100,000 rented next-generation GPUs, Adcock even proclaimed that robot learning likewise obeys a clear Scaling Law:
As long as compute is large enough and pretraining data is rich enough, the motor intelligence of the body will leap exponentially.
Of course, there are also more pragmatic voices, a typical representative being Unitree founder Wang Xingxing. At the World Physical AI Conference some time ago, Wang Xingxing publicly stated that the signs of physical AI being broken through are already extremely clear—one to two years at the fastest, two to three years at the slowest. In his view, within the next two years, the industry will be able to "truly and completely unify foundational large models and robotics technology" at the physical level.
Image source: Unitree
Zhiyuan also takes an optimistic view of the impact of foundational large models on embodied intelligence. Peng Zhihui, co-founder, president, and CTO of Zhiyuan, said that the penetration of general foundational models such as GPT-6 Astra into the embodied field is a major trend, and for a full-stack robotics company like Zhiyuan, insisting on self-developing decision models remains a mandatory choice. Yao Maoqing, co-founder and co-president of Zhiyuan, pointed out more specifically:
As their capabilities improve, large models will increasingly help systems raise the upper limits of long-horizon tasks and planning in System 2, and embodied companies also need to continue investing around System 2 itself.
By contrast, Wang Qian, founder of Independent Variable Robotics, may be more cautious about the capability transfer of general models. At this year's Bund Summit in September, he pointed out that intelligence exists to a large extent in data. The fact that general models can control robots does not prove that intelligence in other domains has naturally transferred to the embodied domain. The fact that OpenAI's and Anthropic's models possess robot manipulation capabilities may be directly related to their having absorbed more physical-world experience.
This is also an important question left by GPT-6 and Fable 5.2: we know they were not specially trained for some test tasks, yet we cannot determine exactly how much robot and physical data they learned during pretraining.
Physical Intelligence co-founder Sergey Levine's judgment leans more toward fusion. He believes that LLMs possess rich general knowledge that can precisely make up for the shortage of robot data, making them a very good starting point for developing robot foundation models. But robots still need to use real manipulation, video, and other physical data to truly transform this knowledge into executable capabilities.
From this perspective, the disagreements among the big names are actually quite clear. Everyone generally acknowledges that general models have brought new capabilities, but there is no consensus on how much independent physical intelligence robots need, or whether these capabilities can be naturally acquired as general models are upgraded.
Moreover, the competition between general models and embodied models may ultimately also involve who controls the data accumulated by robots, and who can continuously improve task success rates in real scenarios. Final thoughts
Six years ago, OpenAI shifted more resources toward large language models because robot data was scarce. Six years later, the development of general models has led it to re-examine the necessity of an independent robotics team.
Today, general multimodal foundation models have already completed the upfront packaging of human society's common sense, visual understanding, and complex logic. Embodied intelligence no longer needs to start learning from an infant's perceptual stage; it stands directly on the shoulders of a super brain with university-level academic ability. But this is also the other side of the coin:
General models have rewritten the starting line of embodied intelligence, but they have not smoothed out the physical gap leading to the finish line.
Even though Fable 5.2 achieved an astonishing leap to a 64.9% success rate in simulation tests, this also means that in the digital simulation world, it still has a 35.1% task failure rate. For robots to truly enter factories, homes, and other complex scenarios, they still need to solve problems such as real-time control, physical interaction, and long-term reliability.
For embodied intelligence companies, the progress of general models is both an opportunity and pressure. On the one hand, they have the opportunity to use stronger foundation models to reduce the difficulty of robots learning new tasks. On the other hand, some intelligence capabilities that originally required specialized development may also gradually become easier to obtain.
This means that in the future, the value of embodied models may increasingly depend on how many physical problems they can solve that general models have not yet solved.
Image source for this article: 123RF licensed image library. Infringement report Return Share Related articles All comments