OpenAI folds its standalone robotics team into core research

OpenAI has dismantled its standalone robotics team. In an exclusive published on November 13, Fortune reported that OpenAI confirmed the reorganization when asked: robotics research staff are moving into the company's main research organization and will report to Chief Research Officer Mark Chen, while the applied robotics and hardware teams are joining the larger Applied division under President Greg Brockman. The company did not say how many people are affected.

An OpenAI spokesperson framed the change as a way to speed up robotics progress rather than a retreat, saying robotics remains "a core part of OpenAI's long-term vision." The company has previously said it is building its own humanoid robot, and Sam Altman's goal of a first public demonstration in 2027 has not been withdrawn.

The shake-up came abruptly. OpenAI had registered booth 208 at CoRL, the robotics research conference held this week in Austin, and the booth appeared on both the conference app and the floor plan. When Fortune's reporter arrived, the space was empty and carried no OpenAI branding; a few former members of the robotics team were milling nearby and still drawing a crowd of people with questions.

The rationale: general models can now drive robots directly

OpenAI told Fortune that earlier this year its researchers believed the best route to better robots was to build on video-generation models such as Sora, since producing realistic video requires an implicit grasp of the physical world. The company has since found that general-purpose frontier models, GPT-6 Astra among them, can be applied to robots directly, which is why it decided to dissolve the separate team and place its people in the groups that build general models.

The report cites a set of experimental results. Researchers hand Astra camera images, the robot's state and a control interface, and let it issue commands itself. In one Nvidia study, an Astra-based agent reached a 74.8% success rate on a simulated manipulation benchmark; the best specialized robotics model on that leaderboard scored 51.0%. Those are results from a simulated environment, not tests on physical robots.

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, told Fortune that he reads the move as OpenAI concluding robotics is "no longer a special problem": something to be built as a downstream application of general intelligence, much as the same frontier model now powers coding, legal and finance tools.

Illustration: robot workbenches move into shared research, reflecting OpenAI team integration and general AI control

2020, 2024, 2026: reading the three reorganizations together

OpenAI had a robotics team early on and shut it down in 2020. Co-founder Wojciech Zaremba later explained that robotics data was limited and there were more efficient paths to AGI. The company then put its resources into training very large language models on huge volumes of text, and two years later it shipped ChatGPT.

In 2024, OpenAI rebuilt a standalone robotics team and announced a partnership with Figure to develop models for humanoid robots. The collaboration lasted only until February 2025, when Figure said it would build its own robot AI in-house.

With this latest move, the team's formal structure is gone but its people remain, and the direction has shifted from building a separate brain for robots to letting the general model double as the brain. What the three reorganizations share is that each followed OpenAI's judgment of its mainline models: when the mainline models offered no hope of being useful to robots, the team was closed, and now that they look sufficient, a separate team is no longer needed.

The reorganization also has a cost backdrop. Last year Altman told the company to cut "side quests"; Sora and other projects have been dropped to free up compute for core businesses more likely to make money. OpenAI reportedly posted an operating loss of about $10.9 billion in its most recent quarter, is pursuing a new funding round, and is preparing for a possible IPO in early 2027. The loss figure comes from media reports and has not been publicly confirmed by OpenAI.

How robotics companies see it

This line of thinking is a challenge to robotics startups, many of which raised billions of dollars on the premise that robots need their own brains. Russ Tedrake, an MIT professor and co-founder and CEO of Walden Robotics, argues that robotics companies hold something AI labs lack: their robots already work in real settings such as factories, and the field data they accumulate helps them adapt to specific jobs, data that becomes more useful the stronger the foundation models get.

Shi also points out that large models are slow and must run in the cloud, which would leave robots permanently tethered to a network, so small specialized models that can run offline on the robot's onboard computer are still needed. Pete Florence, co-founder of robot foundation-model company Generalist, said it is still too early to tell which model will come out on top in robotics.

Sources: Fortune, CocoLoop; OpenAI's confirmation of the reorganization and the reporting lines rest on the company's reply to Fortune; the 74.8% and 51.0% figures are simulated-benchmark results from an Nvidia study; the quarterly loss is as reported by the media.