OpenAI shut down its robotics team again. Its booth at the robot conference sat empty
The company says a general-purpose model can now drive a robot, so it no longer needs a team that only does robots. That is bad news for the startups that raised billions betting it couldn't.
At CoRL, the big academic robotics conference held in Austin last week, the app and the floor plan both said OpenAI was at booth 208. A Fortune reporter went to look. The booth was empty, with no logo in sight.
The explanation came from the company itself. OpenAI has disbanded its stand-alone robotics team and scattered the staff. It confirmed the move when Fortune asked. Robotics researchers now sit in the main research organization under chief research officer Mark Chen. The applied robotics and hardware people have gone to the Applied division, which reports to president Greg Brockman. The change was recent enough that nobody had cancelled the booth.
This is the second time. OpenAI had a robotics group in its early years and wound it down in 2020, when cofounder Wojciech Zaremba blamed a shortage of robot data and said there were faster roads to general intelligence. The company rebuilt the team in 2024 alongside a partnership with the humanoid startup Figure. Figure walked away in February 2025, saying it would build its own robot AI.
A retreat that the company calls an advance
OpenAI's position is that it is doubling down. A spokesperson told Fortune that robotics is still a core part of the long-term vision, and the company says it is developing its own humanoid, with Sam Altman aiming for a first public demonstration in 2027. Breaking up the team, by this account, is the fastest route to general-purpose robots.
The logic goes like this. At the start of the year, OpenAI's researchers assumed the way into robotics ran through models like Sora, the video generator, on the theory that a system able to produce realistic video has absorbed something about how the physical world behaves. Then people started wiring the company's general-purpose model, GPT-6 Astra, directly to robots, and according to Fortune the results surprised them.
The setup in these experiments is almost insultingly simple. Give the model camera images, some information about the robot and an interface for issuing commands. No special training on robot data. In an Nvidia study cited by Fortune, an agent built this way hit a 74.8% success rate on a simulated manipulation benchmark. The best specialized robotics models on that leaderboard had reported 51.0%.
That number comes with a large asterisk: it is a simulation. Nobody's kitchen got cleaned. But it was enough for Guanya Shi, an assistant professor at Carnegie Mellon's Robotics Institute, to read the reorganization as a signal that OpenAI sees robotics as "no longer a special problem." In his reading, the company now treats a robot the way it treats a coding tool or a legal assistant: one more product hanging off the same big model.
Who should be nervous
If OpenAI is right, the awkward position belongs to the robotics startups. As Fortune put it, many of them have raised billions on the premise that robots need brains of their own. A whole category of company exists to build foundation models for robots, and the pitch to investors has been that language models can write a sonnet but can't fold a towel.
The people in that category were, unsurprisingly, not ready to concede. Russ Tedrake, the MIT professor who runs Walden Robotics, argued that robot companies have something the labs lack: machines already working on factory floors and collecting data about specific jobs in specific places. Better general models make that data worth more, he said, because robots learn from it faster.
Pete Florence, cofounder of the robot-model startup Generalist, was more guarded. "There may be different starting points for how to grow intelligence," he told Fortune, adding that it is unclear whether any model that exists today will be the one that ends up dominating robotics.
Shi, for all his reading of the tea leaves, pointed to a practical hole in the OpenAI plan. Big language models are slow and they live in the cloud. A robot that depends on one needs a constant network connection, which is a strange requirement for a machine carrying a pan of hot oil. He still sees a need for small, specialized models that run on the robot's own computer.
There is a habit here. OpenAI closed its first robotics team because robots were too hard and the data too scarce. It has closed the second because, it says, robots have become easy enough to be a feature. Both times the answer was the same: put the people back on the big model.
It is also a convenient conclusion for a company whose entire business is the big model. If one system can run a chatbot, a coding agent and a humanoid, then the money keeps flowing to the same place and the robot specialists become a rounding error. If it can't, OpenAI will have taken apart the one team that knew how to tell.
Fortune's reporter found a few of the former team members standing around booth 208 anyway, with no display and no demo. They still drew a crowd. The question everyone at the conference wanted answered was how much robotics expertise will be worth once general models get good enough, and here were several people whose employer had just given its answer.
Sources: Fortune, Seeking Alpha.
More from the feed
An Eric Trump-backed robot army startup missed payroll. Its CEO borrowed $4.6m for a house
Foundation says it is building humanoid soldiers for the Pentagon. Staff haven't been paid since October 30, federal records show no contracts in its name, and its robot spent 15 minutes failing to load a mortar.
Tesla is finally building Optimus robots. Musk spent the call lowering expectations
The first production-quality humanoids are coming off a Fremont line that used to build cars. Profit fell 43%, the hands are still hard, and outside buyers wait until 2027.
BYD's humanoid robot just leaked through the patent office
China published BYD's 'Embodied AI Robot' design on October 9. The carmaker still hasn't put it on a stage.