OpenAI Disbands Robotics Team as GPT-6 Astra Outperforms Specialized Systems in Simulation
Nvidia data puts Astra at 74.8% on manipulation tasks, nearly 24 points above the top dedicated competitor

OpenAI has disbanded its stand-alone robotics team for the second time, distributing its researchers into the company's main research and applied divisions. The reorganization, confirmed by OpenAI to Fortune and reported Thursday, was driven by an unexpected empirical finding: its frontier general-purpose model GPT-6 Astra can control robots without any dedicated robotics training, and it does so more effectively than the best purpose-built robot AI systems under the same test conditions.
The timing is conspicuous. This week the top academic robotics conference, CoRL — Conference on Robot Learning — is running in Austin. OpenAI had reserved booth 208, which sat empty. Former robotics team members turned up at the booth anyway, drawing a small crowd even without a demo, in what may be an inadvertent illustration of a question now dominating the field: how much specialized robotics expertise remains necessary when a general-purpose model can already outperform dedicated systems in a standard simulation?
A Benchmark That Changed the Strategy
The proximate cause of the restructuring was a study published October 14 by Nvidia's Seattle Robotics Lab, which the company calls Astranomy. Researchers asked a pointed question: could GPT-6 Astra, given camera feeds, robot state information, and a natural-language control interface, run a manipulation robot without any robot-specific fine-tuning? The answer, at least in simulation, was yes — and the margin was larger than expected.
Running against the full RoboLab-120 benchmark, Astra's best configuration — called the Agentic Policy — achieved a 74.8% success rate. The best non-Astra entry on the same leaderboard, a system called Lakeside, reached 51.0% — a gap of nearly 24 percentage points between a general conversational model and systems purpose-built for physical manipulation. For context, the Nvidia study also tested two other Astra configurations: a Direct Policy that bypassed intermediate reasoning and reached 66.3%, and an Orchestrated Policy variant at 66.5%. All three outpaced the best specialized competitor.
The Agentic Policy works by giving Astra a closed-loop interface: it receives visual input, reasons about the state of the robot and the task, issues action commands, and observes the result before the next step. OpenAI told Fortune that this finding — combined with earlier internal work showing that its frontier models could handle robotics tasks without requiring a separate organizational structure — made a standalone robotics team redundant. The company had initially expected progress to build on research from its Sora video generation project, which developed implicit physics knowledge from training on video of physical processes, before Sora was absorbed into a broader world-simulation research group.
What Those Simulation Numbers Actually Mean
The 74.8% figure carries important caveats. The Nvidia study estimates the Agentic Policy costs approximately $36.40 per successful simulated episode, while the model averages 13.37 seconds per inference call. Those economics are untenable for real-world deployment: industrial robots typically operate at 10–30 Hz control loops, requiring decisions in milliseconds, not seconds.
The benchmark itself is entirely simulated. RoboLab-120 consists of 120 pick-and-place tasks in a digital environment, not a physical robot. And a separate academic evaluation on the broader RoboDojo benchmark — 2,100 trials across 42 real-task categories — found that Astra achieves only a 22.48% average success rate, ranking above all 40 public policies but displaying a sharply polarized capability profile: strong on tasks requiring semantic reasoning, weak on tasks requiring precision grasping, bimanual coordination, and dynamic motion.
A third independent evaluation across 12 embodied benchmarks and 8 robot configurations, published in October 2026 by a SenseNova-led research team, found that Astra had "largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence" — a useful distinction that the reorganization announcement largely glosses over.
The Architecture Behind the Gap
Understanding why a general model scores well on manipulation benchmarks requires understanding what those benchmarks actually test. RoboLab-120 is weighted toward tasks that reward planning, scene understanding, and object identification over fine motor precision. A model that reasons fluently about spatial relationships and can map action descriptions onto control commands can handle most of the benchmark's episodes. That maps precisely to Astra's strengths: the model was trained on orders of magnitude more varied semantic data than any specialized robot policy.
GPT-6 Astra was released on September 3, 2026 — just ten weeks before the robotics team was disbanded. The model carries a 1.05-million-token context window and was built primarily for computer use, software engineering, and long-horizon agent work. Its capabilities were not specifically designed for robotics. The fact that it transferred so effectively to manipulation benchmarks — at least in simulation — surprised even researchers familiar with the model's generalist strengths.
The Second Dissolution
This is not OpenAI's first robotics reversal. OpenAI had a robotics team beginning with its founding, targeting general-purpose robots. Co-founder Wojciech Zaremba disbanded that team in October 2020, explaining in a Weights & Biases podcast interview that a lack of training data was holding the program back — robotics had no equivalent of the internet's text corpus, and physical interaction data was expensive and slow to collect.
OpenAI relaunched its robotics unit in May 2024 and confirmed a partnership with Figure AI to develop AI models for humanoid robots. Figure raised $675 million in that round. That partnership ended in February 2025, when Figure AI announced it was developing its own AI models in-house, citing integration challenges and the need for vertical integration in real-world embodied AI.
The current disbanding is the second dissolution in five years. Following the reorganization, robotics research will move under Chief Research Officer Mark Chen, while applied robotics and hardware efforts will join the applied organization led by President Greg Brockman. An OpenAI spokesperson told Fortune that the company remains committed to general-purpose robots, framing the restructuring as evidence of investment rather than retreat. Sam Altman has told Fortune he is targeting 2027 for a first public demonstration of OpenAI's own humanoid robot.
Also worth noting: OpenAI's robotics hardware lead, Caitlin Kalinowski, who joined from Meta in late 2024, resigned in March 2026 — eight months before the team's dissolution — with reports citing concerns about the company's Pentagon deal and lack of guardrails around domestic surveillance and lethal autonomy.
Why Specialized Robotics Companies Are Not Standing Down
The reorganization lands at a moment of genuine uncertainty for the robotics startup ecosystem. Dozens of companies raised large rounds in 2024 and 2025 on the premise that robotics requires a dedicated AI brain — a specialized vision-language-action model trained on robot data, optimized for physical control, and running on hardware close to the robot itself.
Russ Tedrake, co-founder and CEO of Walden Robotics and a professor at MIT, sees a different picture from OpenAI's. Walden, which spun out of Toyota Research Institute in January 2026 and announced a $300 million funding round in July 2026, deployed robots to a Toyota manufacturing plant by February, argues that companies with real-world deployment experience hold a structural advantage AI labs do not possess. Better general models, in Tedrake's view, make field-specific deployment data more valuable rather than less, because a stronger foundation model can extract more signal from real-world physical experience.
That argument has empirical support. Academic benchmarks and industrial deployment do not predict each other reliably: one study found a model achieving 97.65% success on the LIBERO-Spatial simulation benchmark that scored 0% on 68 industrial scenarios in a different evaluation environment. The robotics field has not solved the simulation-to-reality gap, and the benchmark numbers that motivated OpenAI's restructuring do not establish whether Astra's advantages survive physical deployment.
Robotics researchers have noted that general-purpose models like Astra run inference from cloud servers. A robot using Astra as its brain must remain continuously connected to that infrastructure, introducing latency and dependency that purpose-built onboard models avoid. The 13.37-second average inference time of the Agentic Policy configuration makes that constraint tangible: at that speed, a robot cannot respond to a falling object, a moving human, or any dynamic change that arrives faster than its next inference cycle.
What the Reorganization Does Not Settle
OpenAI's decision should not be read as evidence that general models have conclusively solved physical AI. The benchmarks that motivated the restructuring measure specific simulation environments, and the gap between simulation and physical deployment remains robotics' oldest unsolved problem. The 13.37-second average inference time and $36 per episode cost of the Agentic Policy configuration represent a different kind of limitation: a model that cannot operate in real time cannot control a physical robot in a dynamic environment, however well it performs in simulation.
The question the reorganization actually answers is narrower: OpenAI concluded that a dedicated team focused exclusively on robotics is not the right organizational structure when its general-purpose model already shows strong performance on the available benchmarks. That is a different claim from saying general models have replaced specialized systems, or that companies building specialized robotics AI have lost their advantage.
The researchers from OpenAI's former robotics team now distributed across Mark Chen's research division and Greg Brockman's applied organization will help answer those deeper questions — not by building a dedicated robot brain, but by improving a general model that is already, unexpectedly, beginning to work. Whether the same model that runs in cloud-based chat today can run in a physical robot navigating an unstructured environment tomorrow is a question that simulation benchmarks cannot yet answer. The IPO preparations OpenAI is running in parallel — with the company having filed a confidential S-1 with the SEC in June 2026 and targeting public markets in early 2027 — add a financial dimension to the organizational logic: a focused frontier model strategy is easier to pitch to investors than a sprawling hardware program with a 2027 demo target and unresolved deployment economics.



