AI News

OpenAI disbands its independent robotics team, saying GPT-6 Astra can be used directly for robot control

WallstreetcnUpdated Nov 16 20:33

AI summary

OpenAI has disbanded its independently operated robotics team, folding researchers into the main research team and assigning the applied robotics and hardware teams to the Applied organization; the company said the immediate trigger was that general-purpose frontier models such as GPT-6 Astra can be used directly for robot control tasks. The report noted that in an NVIDIA test, an Astra-driven agent achieved a 74.8% success rate on a simulated manipulation benchmark, higher than the 51.0% of the best specialized robotics model on that leaderboard. The restructuring also comes amid OpenAI's cost-cutting, with its operating loss reaching $10.9 billion in the most recent quarter, as it prepares new financing and a possible IPO in early 2027.

Why read this

AI score 78

The original article reviews the organizational destinations and cost background of OpenAI's disbanding of its robotics team, and presents both sides of the debate over whether general-purpose models can replace specialized robotics models.

Topics

Tags

Body · AI translation

OpenAI disbands its standalone robotics team, betting on general-purpose models to directly drive embodied intelligence.

On November 13, Fortune magazine reported that OpenAI has disbanded its independently operated robotics team and reassigned the relevant researchers to other departments. The company said the direct trigger for this move was the unexpected breakthrough of general-purpose frontier models such as GPT-6 Astra.

Researchers found that these models can be applied directly to robot control tasks, eliminating the need to develop dedicated models for robots separately. In a test by Nvidia, an agent powered by Astra achieved a 74.8% success rate on a simulated manipulation benchmark, outperforming the 51.0% reported by the best specialized robotics model on that leaderboard.

The restructuring also comes as OpenAI is under pressure to cut costs. According to reports, the company's operating loss in the most recent quarter reached as high as $10.9 billion, while it is seeking to complete a new round of financing and preparing for a possible IPO in early 2027.

CEO Sam Altman had previously asked the company to abandon "side quests" and concentrate resources on core products.

Where the team went: research and applications merged along two lines

OpenAI confirmed that the robotics research team has been merged into the company's main research team, reporting to Chief Research Officer Mark Chen; the applied robotics and hardware team has been placed under the larger Applied organization, reporting to company President Greg Brockman.

This is already OpenAI's second contraction in the robotics field. The company had an early robotics team, which it cut in 2020.

Co-founder Wojciech Zaremba's reason at the time was the scarcity of robot training data, and that other paths—especially training ultra-large language models on massive amounts of human-written text—could more efficiently advance the goal of developing artificial general intelligence (AGI).

It was precisely this path that ultimately gave rise to ChatGPT at the end of 2022.

OpenAI rebuilt a standalone robotics team in 2024, announcing a partnership with humanoid robotics startup Figure to jointly develop robot AI models. That partnership ended in February 2025, and Figure announced it would develop its own robot AI.

OpenAI said it is still developing its own humanoid robot, and Altman's goal is to complete the first public demonstration in 2027.

The restructuring was so hasty that after disbanding the team, OpenAI still retained a booth registration this week at CoRL, a top academic robotics conference held in Austin. According to Fortune magazine, when the magazine's reporter rushed to Booth 208, the booth was empty and there was no trace of OpenAI's signage.

Core logic: general-purpose models make robotics no longer a "special problem"

In response to Fortune, OpenAI clarified the evolution of the logic behind this strategic shift.

The company said that at the beginning of the year, researchers originally believed that the best path to advance the robotics field was to continue building on the work of video generation models such as Sora. To generate realistic videos, Sora must have some implicit understanding of how the physical world operates.

But afterward, general-purpose frontier models such as GPT-6 Astra demonstrated the ability to be applied directly to robot control, and this discovery prompted the company to change direction.

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, interpreted the adjustment as a signal that OpenAI believes robotics is "no longer a special problem" but a downstream application of general intelligence. Just as the same underlying frontier AI models now simultaneously drive specialized tools in vertical domains such as code, law, and finance.

Shi added that advances in general-purpose models' spatial perception and reasoning capabilities have enabled them to handle tasks that previously required specialized training, such as coping with changes in bottle placement or lighting conditions, and reasoning about and ordering task steps.

This approach poses a potential challenge to the current robotics startup ecosystem. Numerous robotics startups have raised billions of dollars on the premise that "robots need their own brains," and the possibility that general-purpose models can be used directly for robot control fundamentally shakes this narrative.

However, Russ Tedrake, an MIT professor and co-founder and CEO of Walden Robotics, believes robotics companies still have an advantage that AI labs do not possess: their robots have accumulated practical operating experience in real places such as factory floors. This kind of "deployment" data helps robots adapt to specific work scenarios and tasks.

Tedrake pointed out that more powerful foundation models actually make this scenario-specific data more valuable, because robots can learn from it more quickly.

Pete Florence, co-founder of Generalist, took a more cautious attitude. He said that "there may be more than one starting point for cultivating intelligence," and it is currently unclear which existing model will ultimately dominate the robotics field.

In addition, technical constraints remain. Shi pointed out that although large language models are powerful, they run slowly and rely on cloud deployment, meaning robots need to be continuously connected to the internet. He believes that small specialized models that can run independently on a robot's local computer without a continuous network connection still have their value.

Cost pressure: focus on the core, spin off the sidelines

This restructuring is also part of OpenAI's broader cost-control measures.

Altman explicitly asked the company last year to abandon "side quests," and since then OpenAI has successively cut several projects, including the video generation model Sora, in order to reallocate computing resources to core businesses with greater commercial potential.

According to reports, OpenAI's operating loss in the most recent quarter reached $10.9 billion. Under the dual pressure of preparing a new financing round and planning an IPO in early 2027, the urgency of reducing losses continues to rise. Disbanding the standalone robotics team and merging the relevant personnel into mainline research both aligns with the company's strategy of focusing on the core and helps reduce duplicated efforts and improve R&D efficiency.

OpenAI said the reason for relocating robotics researchers to other teams is that the company now believes this is the best way to accelerate the development of general-purpose robots. The company stressed that robotics remains a core component of OpenAI's long-term vision. Risk disclosure and disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.

Source

Wallstreetcn