🔍 Read the full analysis: Could Robot AI Need A Different Approach? OpenAI And Runway Prompt The Question on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
OpenAI has disbanded its stand-alone robotics team and moved researchers into its broader research and applied organizations, telling Fortune that general-purpose models can now be used directly for robotics. Runway has released Praxis-1, a robot-focused model with open weights. Published results suggest general models can perform well on simulated tasks, while specialist models act faster; the tests do not establish which approach works best in real-world robots.
OpenAI has broken up its stand-alone robotics team and moved its researchers into other parts of the company, while Runway has released Praxis-1, a model built specifically for robot control. The contrasting moves, reported by Fortune around the CoRL robotics conference in Austin, spotlight a practical question for the field: whether general-purpose AI can control robots effectively, or whether dedicated models remain necessary for speed and cost.
Fortune reported on November 13 that OpenAI confirmed the reorganization after the outlet found its registered booth at CoRL empty. The robotics research team now reports to Chief Research Officer Mark Chen, while applied robotics and hardware teams have moved to the Applied organization under President Greg Brockman. OpenAI has not published its own announcement about the change.
An OpenAI spokesperson told Fortune that the move reflects the company’s view that frontier models, including GPT-6 Astra, can be applied to robotics directly. The spokesperson described robotics as part of OpenAI’s long-term vision and said the company remains committed to bringing general-purpose robots into the world. Fortune reported that the company had earlier expected to build on video models such as Sora.
Runway’s Praxis-1 takes a specialist approach. The company released the 27-billion-parameter model with weights and fine-tuning code under the Apache 2.0 license, which permits commercial use. On RoboLab-120, Praxis-1 scored 46.2%, according to the source material: the highest score among the listed world action models and open-source robot models, and fifth overall. Runway says it takes about 0.55 to 0.6 seconds per call to generate 1.6 seconds of motion.
Robot control loop · explainer
Who decides the robot’s next move?
Every robot repeats one loop: cameras see, a model decides, the arm moves. The two bets in this story put a different kind of model into the decide step.
Option 1
A general model in the loop
Example: GPT-6 Astra in NVIDIA’s simulation study
18.56 s per model call on average in the direct set-up. Inference was 97.0% of the estimated episode time; the simulation paused while the model was thinking.
Option 2
A specialist robot model in the loop
Example: Runway’s Praxis-1, a world action model with open weights
About 0.55–0.6 s per call on one H100 server, and each call returns 1.6 s of motion. Needs fine-tuning data for each robot.
General model plans the next stepslow, outside the loop
Option 3 · illustrative
Both, in two layers
The idea Runway and the NVIDIA researchers both describe
Runway calls the two approaches “complementary”: an agent plans, a fast policy carries out each step. NVIDIA’s authors say the best way to combine them “remains an open question”.
Sources: NVIDIA Seattle Robotics Lab, “Astranomy” study of GPT-6 Astra, Oct 14, 2026 (simulation, direct set-up); Runway, Praxis-1 technical report, Nov 2026 (one H100 server). Loops and timelines are schematic; the two figures come from different set-ups, not from a head-to-head test.
Graphic: Powered by Thorsten Meyer AI
Speed Separates Robot AI Approaches
The results point to a trade-off rather than a settled winner. In a simulation study by researchers at NVIDIA’s Seattle Robotics Lab, an agent based on GPT-6 Astra reached a 74.8% success rate on RoboLab-120 in an agentic setup. That was above the best trained robot policy on the benchmark leaderboard, according to the researchers. But a direct GPT-6 Astra model call averaged 18.56 seconds, and the simulated environment paused while the model reasoned.
Robot control often requires decisions quickly enough to respond to a changing physical environment. A model that succeeds on more simulated tasks may not be suitable for continuous, real-time action if each decision takes too long or costs too much. The gap matters to companies deciding how to build robot systems, and to customers who may ultimately care about reliability, response time and operating costs as much as benchmark scores.
Runway reported a separate comparison on one of its robots: GPT-6 Astra took 15.6 times longer than Praxis-1 to finish three tasks and cost 51 times more per completed task. That is a company-reported test, not an independent assessment, and the source material does not provide enough detail to generalize it across robots or tasks. It nevertheless illustrates why model speed and inference costs matter alongside task success.
Products Worth Considering
Samsung Galaxy Tab A11+ Kids Edition, 128 GB Android Tablet, Royal Blue
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
DJI Mini 5 Pro Fly More Combo with DJI RC 2
As an affiliate, we earn on qualifying purchases.
Two Models, Different Roles
The NVIDIA researchers tested GPT-6 Astra as a native robot policy: it received camera images, information about the robot and a control interface, then selected movements without robot-specific training. RoboLab-120 is a simulation of 120 tabletop tasks for a single robot arm, so its results describe performance in that test environment rather than general performance in factories, homes or other physical settings.
The study reported three configurations. The direct policy achieved 66.3% success; an agentic policy with persistent history and additional tools reached 74.8%; and an orchestrated setup, in which Astra reviewed movements proposed by a trained robot model, scored 66.5%. In the direct setup, the simulation paused during model reasoning. The study was published by NVIDIA researchers on October 14, 2026, as a CoRL 2026 late-breaking result.
OpenAI’s organizational move is not its first robotics reset. It wound down an earlier stand-alone team in 2020, then rebuilt a team in 2024 and partnered with humanoid maker Figure. Figure said in February 2025 that it would develop its own robot AI, ending that partnership. Fortune also reported cost pressures as business background, but the reason OpenAI gave for the current reorganization was the claimed direct applicability of its general models.
“The two approaches are complementary.”
— Runway
Simulation Results Leave Gaps
The available benchmarks do not resolve whether a general-purpose model or a specialist model is better for real-world robotics. RoboLab-120 is a simulation of tabletop tasks, and the results supplied do not establish how either approach performs across different robots, environments or extended use. The long call time reported in the direct GPT-6 Astra setup also should not be treated as a measure of every possible system configuration.
Runway’s comparison involved one of its robots and three tasks, according to the source material. It is not clear whether the test has been independently replicated, how its cost calculation was made, or whether the same timing gap would hold for other tasks and hardware. OpenAI has not released a detailed statement of its reorganization or published a new robotics evaluation alongside it. The precise responsibilities and future plans for the reassigned teams are also not fully described.
Watch for Physical Robot Tests
The next useful evidence will be comparable tests on physical robots that report task success, response times and costs under the same conditions. Researchers and companies will also need to show whether general models can make decisions quickly enough for practical control, whether specialist models can handle a broader range of tasks, and how the approaches perform when combined.
OpenAI’s spokesperson has said the company remains committed to general-purpose robots, and the source material reports that CEO Sam Altman has targeted a public demonstration in 2027. That target is not a confirmed event date or a guarantee that a demonstration will take place. For now, the reorganization and Praxis-1 release mark competing approaches; the available evidence leaves their real-world performance an open question.
Key Questions
What changed in OpenAI’s robotics organization?
OpenAI disbanded its stand-alone robotics team, moving research staff to the main research organization under Mark Chen and applied robotics and hardware teams to the Applied organization under Greg Brockman, according to Fortune.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter model for robot control. Runway released its weights and fine-tuning code under the Apache 2.0 license, which allows commercial use.
Did GPT-6 Astra outperform Praxis-1?
Not in a direct, like-for-like comparison established by the supplied results. GPT-6 Astra scored 74.8% on RoboLab-120 in an agentic simulation setup, while Praxis-1 scored 46.2% on the same benchmark. Runway separately reported faster completion and lower cost for Praxis-1 on three tasks using one of its robots. The tests measure different setups and do not settle which model is better overall.
Does the research show that general AI can control real robots?
The NVIDIA study tested GPT-6 Astra in a robotics simulation. It does not by itself establish performance across physical robots or real-world environments.
When might OpenAI demonstrate a robot?
The source material says CEO Sam Altman has targeted a public demonstration in 2027. It does not give a confirmed date or further details about what the demonstration would show.
Source: ThorstenMeyerAI.com
