Research
OpenAI Dismantles Its Standalone Robotics Team, Betting GPT-6 Astra Can Control Robots Directly
The lab folded its dedicated robotics organization into two existing divisions on November 13, concluding that its general-purpose frontier model can run robots without a team built around the problem.
Mainstay Communications Team · · 4 min read

In brief
OpenAI confirmed it dissolved its standalone robotics team and split the staff between Chief Research Officer Mark Chen's research organization and President Greg Brockman's Applied division. The company concluded that GPT-6 Astra can be applied to robot control directly, without purpose-built robotics AI sitting beneath it. Early benchmarks back the claim. Researchers in the field are more divided on what it means for the specialized-brain premise that's driven billions in humanoid startup investment.
The Reorganization
OpenAI dissolved its standalone robotics team on November 13, 2026 and confirmed the change to Fortune. Research staff moved under Chief Research Officer Mark Chen. Applied robotics and hardware teams moved to the Applied organization, which reports to President Greg Brockman.
The company told Fortune it views the restructuring as an acceleration of its robotics work. It says robotics remains part of its long-term plans. CEO Sam Altman has separately told Fortune the company's developing its own humanoid robot, with a demonstration targeted for 2027.
The timing had a visible marker. OpenAI had registered booth #208 at the Conference on Robot Learning in Austin, which ran November 10 to 12. Fortune visited the booth and found it empty, with no OpenAI logo.
The Case for Astra
The logic behind the restructuring is that GPT-6 Astra, OpenAI's general frontier model, can operate robots without a separate model trained on robot-specific data.
Researchers at Nvidia gave an Astra-powered agent camera feeds, robot state information, and a control interface. It reached 74.8% on a simulated manipulation benchmark, according to Fortune. The best specialized models on that leaderboard reported 51.0%. The company concluded that a dedicated team was addressing a problem its main research organization could cover.
A second evaluation, published November 14, ran Astra against Ai2's MolmoAct2 on StationeryBench, five bimanual desk tasks on dual-arm YAM robot arms across 200 trials. Astra completed 7 of 100 runs; MolmoAct2 completed 0. Progress scores were 46 and 12 on a 100-point scale. Grading was operator-judged, and MolmoAct2 ran zero-shot outside its training distribution, so the gap has conditions attached.
An independent academic evaluation published at arXiv in September (arXiv:2609.24170) ran Astra through RoboDojo, a 42-task benchmark. Across 2,100 trials, Astra achieved a 22.48% average success rate, ranking above all 40 public entries. GPT-5.5 scored 0.88% on the same benchmark. The paper's authors characterized Astra's profile as "polarized": strong on tasks requiring semantic understanding, weak on tasks requiring precision, dynamic control, or complex bimanual coordination.
A Record of Starts and Stops
It's the second time OpenAI has closed a standalone robotics operation. The first team disbanded in 2020. Co-founder Wojciech Zaremba later said the company lacked sufficient robot training data and concluded that large language models were the faster path to AGI.
OpenAI rebuilt a robotics team in 2024 and partnered with humanoid startup Figure to power Figure's robots. Figure ended that partnership in February 2025, saying it would build its own robot AI. OpenAI's robotics leader Caitlin Kalinowski resigned in March 2026 over objections to the company's Pentagon contracts and their implications for surveillance and lethal autonomy.
The team she left had grown to roughly 100 data collection personnel as of January 2026, according to earlier reporting.
What It Means for the Field
The restructuring applies indirect pressure on a premise behind billions in humanoid funding: that robots need their own specialized intelligence.
Guanya Shi of Carnegie Mellon University reads the move as a signal that robotics is "no longer a special problem," according to Fortune. He adds that cloud-scale models carry latency that makes real-time robot control difficult, which keeps the case for specialized on-device models open in latency-sensitive applications.
Russ Tedrake, an MIT professor and co-founder of Walden Robotics, argues that companies deploying physical robots hold an edge through real-world data accumulated during actual operations. Physical deployment generates observations that synthetic training sets can't replicate at scale.
Pete Florence, co-founder of Generalist, which raised $400 million in June 2026 to build general-purpose robot foundation models, told Fortune it's unclear whether any single model will dominate robotics long-term.
The debate now has a concrete reference point. OpenAI concluded the advantage it once expected specialized robotics AI to hold has closed, at least in simulation. Physical deployment at scale will show whether that conclusion travels to the real world.
Sources
- OpenAI Disbands Its Stand-Alone Robotics Team as Its General-Purpose AI Models Start Running Robots, Fortune
- OpenAI Reassigns Robotics Researchers to Other Teams as General-Purpose Models Might Make Specialized Robotics AI Unnecessary, The Decoder
- An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond, arXiv:2609.24170
- OpenAI Robotics Leader Caitlin Kalinowski Resigns Over Pentagon Surveillance Concerns, Fortune
- Generalist AI Raises $400M at $2B Valuation to Build General Intelligence for the Real World, SiliconAngle
- Walden Robotics Launches with $300 Million to Put General-Purpose Robots to Work Today, The Next Web
Companies mentioned
Mainstay Digital · Spatial and 3D visualization for industrial companies

