What OpenAI Changed in Its Robotics Teams

OpenAI disbanded its stand-alone robotics team and reassigned its staff to research and Applied, according to Fortune's report, and OpenAI confirmed the reorganization to the publication. The robotics researchers joined OpenAI's main research team under Chief Research Officer Mark Chen, while applied robotics and hardware staff joined the larger Applied organization under President Greg Brockman. OpenAI described the reorganization as a way to accelerate work on general-purpose robots. Its spokesperson told Fortune that robotics remains a core part of OpenAI's long-term vision.

The Decoder originally described OpenAI as apparently disbanding its robotics team again. After OpenAI clarified that the work continues across research and product teams, The Decoder corrected its opening sentence to say the team had been reorganized. Neither account reports how many people moved. The reporting lines establish where the work sits. They do not establish how OpenAI will divide future robot projects between those groups.

Fortune also described a visible sign of the change at the Conference on Robot Learning in Austin. OpenAI had registered booth #208, but Fortune found the booth empty, without an OpenAI logo. Some former team members were at the conference. Fortune updated its story on November 13 to clarify their status.

TeamNew homeReports toSource
Robotics research teamOpenAI's main research teamChief Research Officer Mark ChenFortune; The Decoder
Applied robotics and hardware teamsOpenAI's Applied organizationPresident Greg BrockmanFortune; The Decoder
Stand-alone robotics teamDisbanded as a separate team, per Fortune (OpenAI confirmed the reorganization)Not applicableFortune

Why OpenAI Says General Models Can Run Robots

OpenAI says progress with general-purpose models motivated the robotics reorganization, according to Fortune. Its spokesperson said researchers began the year assuming they would build on models such as Sora. As a video model, Sora needs some implicit grasp of the physical world. OpenAI says frontier models including GPT-6 Astra have since shown they can be applied directly to robotics. That is OpenAI's stated reason for changing its team structure. Whether the approach works reliably on physical robots is a separate question from whether OpenAI can organize its researchers around it.

Nvidia's Astranomy report reports results relevant to that rationale. Nvidia evaluated GPT-6 Astra on RoboLab-120, a simulated pick-and-place benchmark using a simulated DROID robot. Its Agentic Policy, which runs Astra as an agent in Codex, succeeded on 74.8% of episodes. Lakeside, the strongest displayed leaderboard reference in Nvidia's report, scored 51.0%. That leaves Astra 23.8 percentage points ahead. Fortune describes that reference as the best specialized robotic model result on the leaderboard. These are benchmark results under Nvidia's stated setup. They are not measured performance for OpenAI's planned humanoid.

The way Astra is used changes the result and its cost. Nvidia reports 66.3% success for Direct Policy and 66.5% for Orchestrated Policy, which delegates actions to a robot foundation model. Cosmos3-Nano-Policy reached 36.8% and pi0.5 reached 28.0% in Nvidia's evaluation. Per successful episode, Nvidia estimates 30.9M tokens and $36.40 for Agentic Policy, 11.0M tokens and $17.72 for Orchestrated Policy, and 160k tokens and $1.96 for Direct Policy. The highest success rate therefore belongs to the costliest Astra setup Nvidia tested.

Nvidia reports 13.37 seconds per inference call for Agentic Policy and 18.56 seconds for Direct Policy. Orchestrated Policy takes 11.50 seconds for an Astra decision plus 2.74 seconds for a Cosmos action request. Nvidia also lists 46.96 milliseconds for pi0.5 and 713.76 milliseconds for Cosmos3-Nano-Policy per action chunk. Those action-chunk figures have a different measurement basis from Astra's inference calls. Nvidia says simulation is paused during inference. A robot's surroundings therefore do not keep changing while the model thinks in this test, so the success rate does not measure its response to a continuously changing scene.

What Separately Published Results Say About Limits

Separately published robot results measure different tasks and cannot be read as a single head-to-head ranking. In a RoboDojo paper by Zhang, Wang, Ouyang and colleagues, GPT-6 Astra averaged 22.48% success across all 42 tasks. It ranked above all 40 public policies the authors compared, while GPT-5.5 reached 0.88% and DeepSeek-Flash reached 1.92%. The authors found Astra stronger on tasks requiring semantic understanding than on precision, dynamic control or complex bimanual coordination.

A separate paper by Zhou, Hou, Fu and colleagues reports 26.92% success on RoboDojo, 76.0% on RoboTwin 2.0 and 11.1% zero-shot success on its LIBERO-GeoMotion benchmark. Its authors also report that zero-shot execution on real hardware is feasible for simple tasks. They identify persistent problems with feasible motion, precise control and physical safety, alongside substantial inference latency. The real-world observation supports a narrow claim about simple tasks. The paper does not establish reliable operation across the precision-heavy tasks it studies.

The Decoder reports another physical-arm comparison on StationeryBench. Across 200 trials covering five desk tasks with dual-arm YAM robot arms, Astra fully completed 7 of 100 tasks and MolmoAct2 completed none. Average progress scores were 46 and 12, respectively. The Decoder separately reports a test in which Astra on a humanoid navigated and cleaned up a kitchen, without giving a benchmark score for that test. These reports show physical-robot activity, but the kitchen account cannot supply a measured success rate or establish repeatability.

The different benchmarks, agent harnesses and episode counts explain why one success rate cannot stand in for another. Nvidia's highest score comes from agentic control of simulated pick-and-place with inference pauses. The RoboDojo and LIBERO-GeoMotion results expose problems with more demanding control. Carnegie Mellon University researcher Guanya Shi told Fortune that general models can handle changes such as a bottle moving or lighting changing. He added that slow, cloud-hosted models leave a need for smaller onboard robot models. Gemini Robotics 2.5 is another general robot model: Google DeepMind's vision-language-action model is rolling out to its robotics partners.

ResultBenchmark and settingReported figureSource
GPT-6 Astra, Agentic PolicyRoboLab-120, simulated DROID robot, pick-and-place74.8% success; Lakeside leaderboard reference 51.0%Nvidia Astranomy report
GPT-6 Astra, Direct PolicyRoboLab-120, same setting66.3% success at 160k tokens ($1.96) per successful episode; 18.56 s per callNvidia Astranomy report
GPT-6 AstraAll 42 RoboDojo tasks22.48% average success, above all 40 public policies; GPT-5.5 0.88%, DeepSeek-Flash 1.92%arXiv 2609.24170
GPT-6 AstraFive benchmarks incl. RoboDojo, RoboTwin 2.0, LIBERO-GeoMotion26.92% RoboDojo; 76.0% RoboTwin 2.0; 11.1% LIBERO-GeoMotionarXiv 2610.28259
GPT-6 AstraStationeryBench, five desk tasks, dual-arm YAM arms, 200 trials7 of 100 tasks fully completed; MolmoAct2 0The Decoder

What It Means for OpenAI's Own Humanoid

Fortune reports that OpenAI has said it is developing its own humanoid robot. Fortune says Chief Executive Officer Sam Altman is targeting a first public demonstration in 2027. The Decoder describes the target as reported. A target is not a confirmed demonstration date. OpenAI's spokesperson told Fortune that the organization remains committed to bringing general-purpose robots into the world. The cited reports publish no specifications, price, release date or hardware details for OpenAI's humanoid.

OpenAI has changed its robotics organization before. Fortune reports that it wound down an early robotics team in 2020. Cofounder Wojciech Zaremba later cited limited robot data and what he saw as a more efficient route to artificial general intelligence (AGI) through other methods. OpenAI rebuilt a stand-alone robotics team in 2024. OpenAI's statement to Fortune says the work has moved into its main research and Applied groups.

Fortune also reports pressure on OpenAI to rein in costs and concentrate on core products. The reporting does not establish a budget for the humanoid or show how the reorganization changes its hardware work. A public demonstration would be an observable milestone if the reported target is met. Published hardware specifications or deployment results would say more about the robot's capabilities than a change in reporting lines.

What It Means for Figure and Other Humanoid Makers

Fortune reports that OpenAI's 2024 partnership with Figure to develop models for humanoid robots ended in February 2025, when Figure said it would develop Figure's own robot AI (artificial intelligence). OpenAI's decision to place robotics researchers with its general research team therefore does not describe a continuing shared model project with Figure. The site's Brett Adcock page covers Figure's founder and the split. The change also does not establish which approach will work better on deployed robots.

Fortune says OpenAI's general-model bet would seem to challenge the premise of robotics startups that raised billions on the idea that robots need dedicated brains. Massachusetts Institute of Technology (MIT) professor Russ Tedrake gave Fortune a reason that robot makers may retain an advantage: their machines gain experience on real factory floors and gather data about specific workplaces. Tedrake said better foundation models can make that site-specific data more valuable because robots learn from it faster. His account concerns deployment experience and the data produced there. It makes no claim that any named maker has won.

There is also a control and operating-cost question. Carnegie Mellon's Guanya Shi told Fortune that large cloud-hosted models are slow and require a constant connection, while smaller specialized models can run on a robot's onboard computer. The Decoder likewise says specialized models may retain advantages in cost, reliability, speed and energy use. Pete Florence, a cofounder of a startup training general-purpose robot models, told Fortune it remains unclear whether any existing model will dominate robotics in the long run. The published results leave room for both general models and dedicated robot models.

Who This Is Not For and What Would Change Our Answer

This account cannot support a robot purchase, deployment or investment decision. The cited sources do not publish OpenAI humanoid pricing or hardware specifications, and the benchmark results do not measure the cost and reliability of that humanoid at a workplace. Readers looking for background on how model-driven robots work can read what embodied AI is. Readers evaluating a particular machine would need evidence about that machine's task, operating conditions and deployment results instead.

A real-hardware benchmark showing reliable precision at acceptable latency would change the assessment of Astra's practical readiness. Published specifications, pricing or deployment results would make OpenAI's own humanoid assessable as a machine rather than a reported plan. A first public demonstration, which Altman is reportedly targeting for 2027, would provide another observable result, though a demonstration alone would not measure workplace reliability. Until such evidence appears, Nvidia's simulated score and the preprints' physical-control limits should remain separate claims.

Our Takeaway

OpenAI says it reorganized its robotics work and has not exited it, according to its statement to Fortune. Published results show that GPT-6 Astra can handle some simulated pick-and-place and semantic tasks. They do not establish reliable precision, acceptable latency or cost-effective operation on real robots. Nvidia's strongest result is 74.8% success in a simulation paused during inference, with an estimated $36.40 per successful agentic episode. Readers tracking humanoid makers should keep that setting attached to the figure when comparing it with the separately published results reporting precision limits.

At The Bot Scout, we read the reports from Fortune, Nvidia, the arXiv authors and The Decoder and compared what each result measures. This is our analysis of the published figures. Real-hardware benchmarks, published OpenAI humanoid specifications or the reported first public demonstration could change the assessment. Until then, the evidence supports OpenAI's reason for investigating general models while leaving the reliability of its planned humanoid and the outcome for dedicated robot makers open.

Bottom Line

Readers examining benchmark claims can read Nvidia's Astranomy report for the simulation conditions, policy types and per-successful-episode costs. Readers following the planned humanoid can watch for the first public demonstration Altman is reportedly targeting in 2027, along with any published hardware details. Those comparing general robot models can read the site's Gemini Robotics 2.5 page. For OpenAI's robotics work, start by checking Nvidia's stated test conditions before using its success rate.

For background on robot foundation models, read what embodied AI is.

FAQs

Did OpenAI shut down its robotics team?

Fortune reports that OpenAI disbanded its stand-alone robotics team, and OpenAI confirmed to Fortune that it reassigned its researchers and applied staff. OpenAI said robotics remains part of its long-term vision. After OpenAI clarified that robotics work continues across other teams, The Decoder corrected its opening description to call the change a reorganization.

Who runs OpenAI's robotics work now?

OpenAI's robotics researchers moved to its main research team under Chief Research Officer Mark Chen, according to Fortune and The Decoder. Its applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. The reports establish those reporting lines but do not say how many staff moved or assign every future robot project to a particular group.

Is OpenAI building a humanoid robot?

Fortune reports that OpenAI has said it is developing its own humanoid robot. Chief Executive Officer Sam Altman is targeting a first public demonstration in 2027, according to Fortune. The Decoder calls that target reported. Neither account provides the robot's hardware specifications, price or release date. The reported demonstration target is not a published commitment to deliver a robot in 2027.

Can GPT-6 Astra control a real robot?

Yes, for some simple tasks, according to published reports, but reliable real-world control remains unproven. The authors of an arXiv paper report feasible zero-shot execution on simple real-world tasks alongside precision, safety and latency deficits. Nvidia's 74.8% result comes from simulation. The Decoder describes a humanoid kitchen-cleanup test without publishing a benchmark score for it.

What happened between OpenAI and Figure?

OpenAI and Figure announced a humanoid robot model partnership in 2024, according to Fortune. It ended in February 2025 when Figure said it would develop its own robot artificial intelligence. The end of that partnership does not establish which model approach will perform better on deployed robots. The site's Brett Adcock page covers Figure's founder and the split.

Primary Sources