The headline sounds like a retreat: OpenAI has disbanded its stand-alone robotics team. The explanation points in the opposite direction. The company thinks its general-purpose models can do more of the thinking that used to require a separate robotics effort.
That is a meaningful strategic bet. It is not an announcement that a chatbot can now be dropped into a factory and trusted with the machinery.
The interesting story sits between those two interpretations: general models are becoming more useful at reasoning about physical tasks, while the evidence still shows a considerable gap between an impressive experiment and reliable work.
A reorganization, not an exit from robotics
Fortune reported on November 13 that OpenAI confirmed the reassignment of its robotics researchers into the main research organization under chief research officer Mark Chen. Applied robotics and hardware teams moved into the broader Applied organization led by Greg Brockman.[1]
OpenAI told Fortune that robotics remains part of its long-term vision and presented the change as a way to accelerate it. The company's explanation was that frontier models, including GPT-6 Astra, had shown they could be applied to robotics more directly than its earlier approach anticipated.[1]
That account deserves careful attribution. It comes from OpenAI's response to Fortune, not a separately verified product release. Reorganizing researchers does not establish that their hardest engineering problems have been solved.
The underlying idea is nevertheless important. Instead of treating language, vision, coding and physical reasoning as completely separate disciplines, a company can try to reuse a general model's ability to interpret a scene, break a goal into steps and recover from an unsuccessful action.
Whether that becomes an economical and dependable robot is a different question.
The 74.8 percent result needs its full label
NVIDIA's Astranomy research report evaluates GPT-6 Astra on RoboLab-120, a benchmark of 120 simulated manipulation tasks. Its strongest reported agentic configuration achieved 74.8 percent success, compared with a 51.0 percent leaderboard reference for Lakeside.[2]
The result is not a single model answering one prompt and directly driving motors. The agentic configuration retains episode history through a persistent Codex session. The report distinguishes it from a direct-policy configuration, which scored 66.3 percent, and an orchestrated configuration, which scored 66.5 percent.[2]
Most importantly, the simulation pauses while the model is thinking. The evaluation therefore does not require the system to keep up with the physical world in real time. The agentic configuration's reported average inference call took about 13 seconds.[2]
That is not a footnote. A simulator can wait indefinitely while a robot decides what to do next. A person, moving conveyor or falling object cannot be assumed to wait with it.
The displayed 51.0 percent comparison is also a leaderboard reference, not evidence that every specialized robotics system was freshly tested under an identical integration. The research is encouraging without needing that stronger claim.
Real hardware tells a more modest story
StationeryBench provides a useful counterweight. Its published experiment compared Astra with a released MolmoAct2 checkpoint on five tabletop tasks using bimanual YAM robot arms. Each model received 100 trials: 20 per task.[3]
The tasks were deliberately ordinary: uncap a marker, move paperclips into a bowl, hand over a ruler, remove a sticky-note pad and retrieve an eraser from a box before closing the lid. Ordinary objects do not make the manipulation easy.
The headline result was seven completed tasks out of 100 trials for Astra, versus zero for the tested MolmoAct2 configuration. Mean progress scores were 46 and 12, respectively. Those progress scores came from human-judged milestones; they were not completion percentages.[3]
| Experiment | Reported result | What it does not establish |
|---|---|---|
| Astranomy agentic policy in simulation | 74.8% task success | Real-time physical performance at that rate |
| StationeryBench on physical arms | 7 completed trials out of 100 for Astra | Reliable operation across a normal workday |
| StationeryBench progress scoring | 46 versus 12 points | 46% versus 12% completed tasks |
The physical comparison also has important limits. The models used different action interfaces and time budgets. MolmoAct2 was evaluated zero-shot without task-specific fine-tuning. The researchers disclose that not all trials used the same physical rig, and graders knew which model they were watching.[3]
None of that makes the experiment worthless. It makes the scope of the conclusion clearer: a general model showed promising capabilities in this setup, but one specialized checkpoint's weak result cannot settle the entire generalist-versus-specialist debate.
The controller did not disappear
The most useful engineering detail in the Astranomy report is less dramatic than its headline score. The model proposes actions, while the host validates and executes them. Motion bounds and inverse kinematics remain part of the system.[2]
That distinction matters. A model can contribute perception, planning or error recovery without replacing the mechanisms that turn a request into bounded physical movement.
For an industrial buyer, a better question than whether a robot has a general-purpose brain is whether the complete application behaves safely and predictably. That includes calibration, sensors, end effectors, controllers, communication failures and the people who work nearby.
OSHA's industrial-robot technical guidance describes hazards from unexpected movement, component failures and changes in programming or peripheral equipment. It emphasizes the robot application and its environment, not simply the intelligence of one software component.[5]
This is U.S. technical guidance, not a claim that one document determines compliance in every country. The general engineering lesson travels well: a more capable model does not remove the need for a task-specific risk assessment and appropriate safeguards.
Benchmark settings are part of the result
There is another small but revealing detail. The current StationeryBench repository explicitly distinguishes its packaged evaluation defaults from the published experiment: the package uses a different default duration and a binary operator verdict, while the paper's reported agent condition used its own horizon and video-graded stages.[4]
A buyer or researcher who reproduces the package with its defaults is therefore not automatically rerunning the exact published condition.
This is a wider AI problem. A score is a property of a model, harness, task distribution, action interface and evaluation procedure together. Changing the surrounding system can change what the number means, even when the model name stays the same.
What would make this useful outside the lab?
For businesses, the credible opportunity is not necessarily an all-purpose humanoid arriving next year. It is a constrained task becoming easier to teach, more tolerant of variation or less dependent on hand-written exception handling.
The sensible deployment questions are concrete:
- Can it complete the actual task repeatedly, rather than make partial progress?
- What happens when an object moves, a camera is blocked or a network connection fails?
- How long does a successful cycle take, including thinking and recovery?
- How often does a person need to intervene, and what does that intervention cost?
- Which controls prevent an unsuitable model proposal from becoming unsafe motion?
These are purchasing and evaluation criteria, not claims that OpenAI has already answered them. A useful trial should compare the proposed system with the existing process on the same work, including failed attempts and supervision.
OpenAI's reset makes sense as a bet on where more of a robot's reasoning may come from. The experiments support taking that bet seriously. They do not show that physical-world reliability has become a solved problem.
The robot's reasoning may become more general. The evidence for trusting its work still has to be specific.
Sources checked November 14, 2026. This is analysis of published reporting and experiments, not an Intueo hardware test or a product endorsement.
References
- [1]Fortune: OpenAI disbands its stand-alone robotics team, November 13, 2026
Original reporting with OpenAI's confirmation and explanation of the reorganization.
- [2]NVIDIA Research: Astranomy
Primary report describing agent configurations, paused simulation, success rates, latency and the host-controlled action interface.
- [3]RoboCurve: StationeryBench experiment report
Primary physical-robot results, trial conditions, progress scoring and disclosed limitations; this is not an OpenAI corporate announcement.
- [4]StationeryBench source repository
Documents the benchmark implementation and differences between packaged defaults and the published experiment.
- [5]OSHA Technical Manual: Industrial Robot Systems and Safety
Technical context for application-level hazards and safeguards, not a certification of any model or installation.


