OpenAI folds its robotics team into its main research and applied groups
OpenAI says robotics work continues and that models like GPT-6 Astra can now handle robot tasks directly. Outside experts disagree on what that means for specialised robot models.
Key takeaways
- OpenAI has moved its robotics researchers under chief research officer Mark Chen and its applied robotics and hardware staff into Greg Brockman's Applied organization.
- OpenAI confirmed the standalone team has ended but says robotics work continues and the move is meant to speed it up.
- OpenAI argues general models like GPT-6 Astra are now able to do robot tasks, citing a 74.8% success rate in an Nvidia simulation.
- On real robot arms, Astra fully completed 7 of 100 desk tasks in a separate benchmark.
- Outside researchers disagree on whether specialised robot models will remain necessary.
OpenAI has ended its standalone robotics team and moved its people into other parts of the company, according to The Decoder, BigGo Finance and AI Weekly, which cite Fortune's reporting. OpenAI confirmed the change on Thursday, Nov. 13, after questions from Fortune, BigGo Finance reports. Its robotics researchers now sit under Mark Chen, OpenAI's chief research officer, and staff working on applied robotics and hardware have moved to the Applied group run by Greg Brockman, the company's president, the outlets report.
OpenAI confirmed the standalone team no longer exists but rejects the idea that it is leaving robotics, according to BigGo Finance and The Decoder. The Decoder first wrote that OpenAI was disbanding its robotics team, then changed its opening after OpenAI said the team had been reorganized to bring it closer to research and product teams, and that its robotics work goes on. A company spokesperson said robotics stays central to OpenAI's long-term plans and described the change as a way to speed up the work rather than step back from it, according to BigGo Finance and AI Weekly.
Why did OpenAI make the change?
OpenAI says frontier models such as GPT-6 Astra have proved able to take on robot tasks directly, reducing the need for a separate team building specialised robot models, according to BigGo Finance; The Decoder says OpenAI apparently holds that view. The company cites an Nvidia simulation study, BigGo Finance and AI Weekly report, in which an agent built on Astra succeeded in 74.8% of simulated manipulation tasks, against 51% for the best specialised models. The Decoder also reports the result.
That result has limits. The Decoder notes the test used a simulated robot, not a physical one. BigGo Finance reports that the 74.8% figure came from a version of Astra that could use tools and keep a record of the task, that the model acting on its own scored lower but still ahead of the specialised models, and that each task was run only five times. In the same study, each command took an average of 18.56 seconds while the simulation was paused, and the agent's estimated API cost was $36.40 per successful task, BigGo Finance and AI Weekly report. On real dual-arm robots, Astra fully completed 7 of 100 desk tasks in a separate benchmark, while Ai2's MolmoAct2 completed none, according to The Decoder and AI Weekly.
The change came at short notice: OpenAI's registered booth at CoRL, the robot learning conference held in Austin that week, stood empty, The Decoder, BigGo Finance and AI Weekly report.
What do outside experts say?
Guanya Shi of Carnegie Mellon University's Robotics Institute said the move suggests OpenAI no longer sees robotics as a special problem but as one more application of general intelligence, according to BigGo Finance and AI Weekly. He also said large models are slow and usually run in the cloud, so smaller specialised models that run on a robot's own computer are still needed, BigGo Finance reports.
Russ Tedrake, an MIT professor who runs Walden Robotics, said robot makers keep an advantage because their machines are already gaining experience in real workplaces, according to the same outlets. Pete Florence, cofounder of the robot AI startup Generalist, said it is not clear that any of today's models will come to dominate robotics, AI Weekly and BigGo Finance report.
What is OpenAI's robotics history?
OpenAI shut down its first robotics team in 2020, when cofounder Wojciech Zaremba cited a lack of robot data, and rebuilt a team in 2024, The Decoder, BigGo Finance and AI Weekly report. A partnership with humanoid maker Figure ended in February 2025. Chief executive Sam Altman is aiming for a first public demonstration of OpenAI's own humanoid robot in 2027, according to BigGo Finance and AI Weekly; The Decoder says he is reportedly targeting a robot demo that year. The Decoder adds that OpenAI is under financial pressure to concentrate on its main AI products.
Questions readers are asking
Has OpenAI stopped working on robots?
OpenAI says no. It says it reorganized the team so robotics research sits in its main research group and applied and hardware work in its Applied organization, and that robotics remains part of its long-term plans.
Why did OpenAI reorganize the team?
It says its frontier models, GPT-6 Astra among them, can be used directly for robot tasks, citing an Nvidia simulation study.
How well does GPT-6 Astra control real robots?
In one benchmark on real dual-arm robot arms, it fully completed 7 of 100 desk tasks, while Ai2's MolmoAct2 completed none. The 74.8% figure OpenAI cites comes from a simulation.
Sources
- The Decoder: OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary
- BigGo Finance: OpenAI Disbands Standalone Robotics Unit, Folds Staff Into Core Research and Applied Teams
- AI Weekly: OpenAI folds robotics team into main research and applied orgs
- BigGo Finance: OpenAI Scraps Standalone Robotics Team as General-Purpose Models "Take Over" the Embodied AI Race
- #openai
- #robotics
- #gpt-6-astra
- #embodied-ai
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Agents3 min
OpenAI launches ChatGPT Wallet so its dots agents can buy for users
The payment feature starts with US ChatGPT Pro subscribers. Users approve each purchase with a biometric check, and agents get a payment token, not the real card number.
Robotics2 min
Tesla invites select buyers to order a Cybercab for personal use
The driverless Cybercab costs $29,990 plus a required $299-a-month autonomy subscription and only runs inside Tesla's robotaxi service areas.
Agents3 min
OpenAI's ChatGPT Wallet lets its dots agents buy things for users
Built with American Express, Mastercard and Visa, it starts with some US ChatGPT Pro users. Each purchase needs user approval, and OpenAI says agents never see card details.
Chips & Infra3 min
TSMC board approves up to $500 million each for Anthropic and OpenAI
The chipmaker's board cleared the stakes on Nov. 10, along with about $40.9 billion in capital appropriations. The filing does not say when either purchase will happen.