OpenAI disbanded its autonomous robotics team, moving its researchers to other teams.
The company confirmed the reorganization at the request of Fortune. A company spokesperson said robotics remained a critical part of OpenAI’s long-term vision and presented the reorganization as a sign of the company’s commitment to robotics, rather than a retreat.
OpenAI said it reassigned its robotics research team to the company’s core research team, reporting to Chief Research Officer Mark Chen, while the applied robotics and hardware teams were reassigned to the company’s larger applied organization, reporting to OpenAI President Greg Brockman.
The change happened recently enough that OpenAI has already registered a booth at CoRL, a premier academic robotics conference in Austin, this week. Although the conferencing app and floor plan marked OpenAI’s place, when Fortune visited booth #208, it was empty and the OpenAI logo was nowhere to be found.
OpenAI had a robotics team in its early days, but disbanded it in 2020, with co-founder Wojciech Zaremba later citing limited robotics data and a more efficient path to artificial general intelligence (AGI) – the human-like AI that has been OpenAI’s goal since its inception – through other methods. In particular, OpenAI has increasingly focused on training ultra-large language models on large datasets of texts written by humans. It’s this approach that ultimately gave birth to the company’s revolutionary ChatGPT product in late 2022.
OpenAI has rebuilt an autonomous robotics team in 2024, announcing a partnership with startup Figure to develop AI models for humanoid robots. This partnership ended in February 2025, when Figure announced that it would develop its own robotic AI. OpenAI said it is currently developing its own humanoid robot, with CEO Sam Altman aiming for a first public demonstration in 2027.
The company spokesperson said that despite the reorganization, OpenAI remains committed to bringing general-purpose robots to the world.
This reorganization reflects a bet that is gaining traction in the AI industry: General-purpose models designed for language, code, and vision can also be used for “embodied AI,” where AI models control devices that can act in the physical world, such as robots, driverless cars, or drones. This would avoid having to have separate robotics teams building specialized AI models just to be the digital “brains” of the robots.
Researchers saw surprising performance when connecting the GPT-6 Astra to robots, directly using the general-purpose AI model to perform certain robotic tasks.
Guanya Shi, assistant professor at the Robotics Institute at Carnegie Mellon University, who had previously heard about the OpenAI revamp, said: Fortune that he would interpret this decision as a signal that OpenAI considers robotics to be “no longer a particular problem”. Instead, he says, they likely view robotics as a challenge that can be solved as a downstream application of general intelligence — in the same way that the same underlying cutting-edge AI models now power specialized coding tools or tools for legal or finance work.
OpenAI seemed to confirm this view in its response to Fortune. Earlier this year, the company’s researchers believed the best way to advance robotics was to build on their work on AI models such as Sora, a video generation model that, to create realistic videos, needed to have some implicit understanding of how the physical world worked.
But since then, the company spokesperson said, the company’s general-purpose frontier models, including the GPT-6 Astra, have shown they can be directly applied to robotics. This discovery motivated OpenAI to disband the autonomous robotics team and move affiliated researchers to teams working on more general AI models.
The reorganization also comes at a time when OpenAI is under pressure to rein in costs and show a greater focus on a few core products. Last year, Altman told the company it needed to abandon “side quests,” and the company abandoned some efforts, including Sora, in order to reallocate computing resources to more core and potentially profitable activities. The company is under pressure to stem its hefty operating losses – which reportedly totaled $10.9 billion in the most recent quarter – as it seeks to close a new funding round and prepares for a possible IPO in early 2027.
This line of thinking – that general AI models can be applied directly to controlling robots without the need to create specialized models with “spatial intelligence” – would seem to challenge the assumptions underlying many of today’s robotics startups, which have raised billions on the premise that robots need their own brains.
But Russ Tedrake, an MIT professor and co-founder and CEO of Walden Robotics, said robotics companies still have an advantage that AI labs don’t: Their robots already gain work experience in real-world locations, like factories. Such a “deployment” experience – and the data collected by these robots – helps robots adapt to specific workplaces and tasks. Better fundamental models make this site-specific data more valuable because robots can learn from it more quickly, Tedrake said.
Recent experiments have tested how much work a general-purpose model can do without specialized training on robotic data. The researchers provided Astra with camera images, information about the robot and a control interface allowing the model to command the robot to perform certain actions. In an Nvidia study, an Astra-powered agent achieved a 74.8% success rate on a simulated handling test, where the model controls a simulated robot in a digital simulation, as opposed to a real robot, compared to 51.0% reported by the top specialized robotic models in this ranking.
Shi said advances in general models’ spatial and reasoning abilities are already allowing them to handle tasks that previously required dedicated training. They can cope when a bottle is in a different position or the lighting changes, and reason about the sequence of a task or which step went wrong, he said.
At the same time, large language models are powerful but slow and must be hosted in the cloud, which means the robot or device needs a constant network connection. Shi still sees the need for smaller, specialized models that can run on a robot’s onboard computer without needing a constant internet connection.
“There can be different starting points for developing intelligence,” said Pete Florence, co-founder of Generalist, a startup that trains general-purpose AI models for robots. Fortune. He added that it was unclear whether an existing model would dominate robotics in the long term.
For its part, OpenAI said it is reassigning its robotics researchers to other teams because it now believes that is the best way to accelerate progress toward building general-purpose robots.
Back at CoRL booth #208, a few former members of the once unified and autonomous OpenAI robotics team gathered. They still attracted a crowd, despite the lack of an exhibition or demonstration. After all, one of the most burning questions at the conference was how valuable robotics expertise would remain as general-purpose models became more capable — and the experience of these former OpenAI robotics researchers, now sent to other teams, offered an answer.
Gn bussni
