OpenAI folds its robotics team into core research, betting GPT-6 Astra is the robot's brain
OpenAI has disbanded its stand-alone robotics team, Fortune reported, moving researchers under Mark Chen and hardware staff under Greg Brockman. The company says its frontier models can run robots directly, which puts it against the startups building separate robot models.
OpenAI has broken up its stand-alone robotics team, Fortune reported on Nov. 13. Robotics researchers now sit in the main research organization under chief research officer Mark Chen, and the applied robotics and hardware teams have moved into the Applied organization that president Greg Brockman runs. A spokesperson told Fortune that "robotics remains a core part of OpenAI's long-term vision," and the company says the target for a first public humanoid demo is still 2027.
The move is a bet that a robot's brain is the general model itself. OpenAI told Fortune its frontier models, including GPT-6 Astra, have shown they can be applied directly to robots, so robotics no longer needs its own organization. The public evidence offered so far is a single result: in an Nvidia study Fortune cited, an Astra-powered agent reached a 74.8% success rate on a simulated manipulation benchmark, against 51.0% for the best specialized robot models on the same leaderboard. That is one simulation, run by someone else.
The hardware team now reports to the product side
A month ago the hiring pointed the other way. In October, OpenAI's careers page listed 27 robotics openings, Engadget counted, including custom actuators, motors, thermal systems and supply chain. That work now sits under Brockman's Applied group, next to products, while the research that decides what the robot can do sits with the people training GPT. When Robrief asked on Oct. 18 whether the 2027 demo would show a robot OpenAI built or an OpenAI model steering someone else's hardware, this reorganization tilts the answer toward the model.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.
This is OpenAI's third structure for robotics. It wound down its first team in 2020, rebuilt a stand-alone group in 2024 alongside a partnership with Figure, and lost that partner in February 2025 when Figure said it would build its own robot AI. Fortune noted that OpenAI had registered booth #208 at CoRL in Austin this week, and the booth was empty when its reporter visited.
The specialists' answer is deployment data
Guanya Shi of Carnegie Mellon told Fortune robotics is "no longer a special problem," while still seeing a place for smaller models that run on the robot without a constant connection. Pete Florence, co-founder of Generalist, which trains robot foundation models, said "there may be different starting points for how to grow intelligence." Russ Tedrake, the MIT professor who runs Walden Robotics, argued that robot companies keep an edge because their machines collect real-world deployment data. A simulated benchmark does not measure that.
Sources Sources verified
- #OpenAI
- #GPT-6 Astra
- #robotics team
- #frontier models
- #reorganization
- #CoRL 2026
- #Generalist
- #Walden Robotics
Share this briefing
Know someone who follows robotics? Send it their way.
Related briefings
More in Embodied AI & Foundation ModelsNVIDIA invests in RLWRLD as the Korean startup releases DexBench, a 15-task test for humanoid hands built on NVIDIA's simulator
NVIDIA said on Nov. 9 it will invest directly in Seoul's RLWRLD, which the same day showed DexBench, a dexterity benchmark that runs on Isaac Lab-Arena. The stake buys NVIDIA a say in where robot hands get graded, not just another model maker.
LG targets 500,000 hours of robot training data and puts home humanoids last, around 2030
On its Q3 earnings call, LG Electronics said it is building about 500,000 hours of robot training data with NVIDIA and will expand into home humanoids around 2030, after manufacturing and logistics. The home, where its CLOiD robot debuted, now comes after parts, data and factory robots.
NVIDIA's GR00T 2 learns mostly from human video, then needs five robot demos to hit 95%
NVIDIA released GR00T 2 at CoRL, a world-action model pretrained on about 436,000 hours of data, 92% of it human video. On 50 tasks its target robot never practiced, it succeeds 40% of the time zero-shot and 95% after five demonstrations, which moves the costly part of robot data to the last few minutes per task.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.