AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Do Robots Need Their Own AI? OpenAI Reshapes Robotics As Runway Launches Praxis-1 on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

OpenAI has disbanded its stand-alone robotics team, moving researchers into its main research group and applied robotics and hardware staff into its Applied organization. Runway released Praxis-1, an open-weight robot model that its own tests say acts faster and costs less per task than GPT-6 Astra; benchmark results suggest general models may handle more simulated tasks while specialist models operate faster.

OpenAI has broken up its stand-alone robotics team, moving researchers into its main research group and applied robotics and hardware teams into its Applied organization, while Runway released Praxis-1, an open-weight model built for robot control. The moves frame a practical question for the field: whether general-purpose AI can control robots on its own, or whether dedicated models are still needed for fast, real-world action.

Fortune reported on November 13 that OpenAI confirmed the restructuring after the outlet asked about an empty company booth at the CoRL robotics conference in Austin. Robotics researchers now report to Chief Research Officer Mark Chen; applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. OpenAI has not published its own announcement, and its spokesperson characterized the shift as a commitment to robotics rather than a retreat.

The spokesperson told Fortune that OpenAI’s view had changed since the start of the year, when researchers expected to build on video models such as Sora. The company now believes its general-purpose frontier models, including GPT-6 Astra, can be applied to robotics directly. That is the company’s explanation for the reorganization; the available benchmark results measure performance in simulation and do not establish how well the approach works on physical robots.

Runway’s Praxis-1 is a 27-billion-parameter “world action model”, released with weights and fine-tuning code under the Apache 2.0 license. On RoboLab-120, it scored 46.2%, the highest among the world action models and open-source robot models listed in the benchmark, and fifth overall. Runway says Praxis-1 takes about 0.55 to 0.6 seconds per call to generate 1.6 seconds of motion. Those figures are not directly interchangeable with results from every other test setup.

At a glance
reportWhen: OpenAI’s reorganization was confirmed b…
The developmentOpenAI reorganized its robotics teams as Runway released the specialist robot model Praxis-1, renewing debate over whether robots need dedicated AI.

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”.

The trade-off so far: general models solve more varied tasks in simulation; specialist robot models are the ones fast enough to act in real time.

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 Versus Simulated Task Success

The split matters because robot control has different demands from answering a question or generating an image. A model that can solve a task in principle may still be unsuitable if it takes too long to decide what the robot should do next. In NVIDIA researchers’ simulation study, GPT-6 Astra’s strongest setup reached 74.8% success on RoboLab-120, but the direct setup averaged 18.56 seconds per model call. The simulation paused while the model processed each call, so that measure does not represent uninterrupted physical operation.

Runway reports a different kind of advantage. In its own test on one of its robots, it says GPT-6 Astra took 15.6 times longer than Praxis-1 to finish three tasks and cost 51 times more per completed task. These are company-reported results from a limited test, not an independent head-to-head evaluation across a broad range of robots and tasks. Together, the figures show why task success alone may not determine which systems are useful: latency, operating cost and reliability also matter.

The approaches need not be mutually exclusive. A general model could interpret instructions or plan a sequence of tasks, while a faster specialist model handles immediate movements. Runway describes the systems as “complementary”. That is a possible division of work, not a confirmed architecture or a settled industry standard.

Products Worth Considering

LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor

LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor

As an affiliate, we earn on qualifying purchases.

From Robotics Team to Model Integration

This is the second time OpenAI has wound down a stand-alone robotics team. The company closed its first team in 2020. It rebuilt a robotics group in 2024 and partnered with humanoid maker Figure; Figure said in February 2025 that it would develop its own robot AI, ending that partnership. OpenAI has also said it is developing a humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027.

The evidence cited for the current shift includes a simulation study published October 14, 2026, by researchers at NVIDIA’s Seattle Robotics Lab as a CoRL 2026 late-breaking result. It tested GPT-6 Astra on RoboLab-120, a simulation of 120 tabletop tasks for one robot arm. The model received camera images, information about the robot and a control interface, and produced actions without robot-specific training.

The study reported three configurations: a direct policy scored 66.3%; an agentic policy using a persistent Codex session and additional tools scored 74.8%; and an orchestrated setup, in which Astra reviewed proposals from Cosmos3-Nano-Policy, scored 66.5%. The agentic result exceeded the benchmark’s best listed policy, Lakeside, by 23.8 percentage points, according to the researchers. These are simulated results, not proof of equivalent performance in physical environments.

“The two approaches are “complementary.””

— Runway

Real-World Performance Still Unknown

It is not yet clear how the simulation results translate to physical robots. RoboLab-120 covers tabletop tasks for a single robot arm; it does not settle performance across different hardware, environments or longer tasks. The study’s direct GPT-6 Astra setup averaged 18.56 seconds per call, but the reported test paused the simulation during inference, leaving real-time performance under continuous physical operation unresolved.

Runway’s speed and cost comparison is the company’s own test on one of its robots and three tasks. The supplied information does not provide enough detail to judge how broadly the result applies or whether an independent evaluation would reproduce it. OpenAI has not posted a public statement with additional detail on the reorganization, and it has not publicly described how its general models will be deployed in robot-control systems.

The available results also use different setups and measures. Praxis-1’s 46.2% score and GPT-6 Astra’s 74.8% score come from different model categories and configurations on the same benchmark, but the figures alone do not show which system is preferable for a particular robot or task. The cost, safety and reliability of deploying either approach outside the reported tests remain unconfirmed.

Evidence to Watch in 2027

The next useful evidence will be independent testing on physical robots, with comparable tasks, hardware, timing and cost measures. Further reporting or a direct OpenAI statement may clarify which teams and projects moved, how the company plans to integrate robotics work with general-model research, and whether the new structure changes development plans.

OpenAI’s stated target for a public humanoid-robot demonstration in 2027 is another milestone to watch, though no specific date or demonstration details are provided here. Runway’s release makes Praxis-1’s weights and fine-tuning code available for commercial use under Apache 2.0, allowing others to test and adapt the model. Whether those efforts favor one model type—or combine general planning with specialist control—remains an open question.

Key Questions

What changed at OpenAI?

According to Fortune, OpenAI confirmed that it disbanded its stand-alone robotics team. Robotics researchers moved under Chief Research Officer Mark Chen, while applied robotics and hardware teams moved into the Applied organization under Greg Brockman.

What is Runway’s Praxis-1?

Praxis-1 is a 27-billion-parameter robot model described as a world action model. Runway released its weights and fine-tuning code under the Apache 2.0 license, which allows commercial use under that license.

Did GPT-6 Astra outperform Praxis-1?

The reported figures do not establish a definitive overall winner. GPT-6 Astra’s agentic setup scored 74.8% on RoboLab-120 simulation, while Praxis-1 scored 46.2%; Runway separately reports that Praxis-1 acted faster and cost less in a test of three tasks on one of its robots. The results reflect different setups and should not be treated as a single head-to-head verdict.

Why might robots use a specialist AI model?

Robot control often requires rapid decisions. In the cited NVIDIA study, direct GPT-6 Astra calls averaged 18.56 seconds, while Runway says Praxis-1 generated 1.6 seconds of motion in about 0.55 to 0.6 seconds per call. Those measurements use different test setups, but they illustrate why response time can matter alongside task success.

Is OpenAI leaving robotics?

OpenAI says no. Its spokesperson described robotics as part of the company’s long-term vision and said it remains committed to general-purpose robots. The company has said it is developing a humanoid robot and has set a public demonstration target for 2027, but further details of the reorganization have not been published by OpenAI.

Source: ThorstenMeyerAI.com

You May Also Like

RHEO on the Web: Find Your Flow

Discover RHEO’s web version, a private, instant, browser-based fluid simulation for relaxation, breathing, and creative exploration—no sign-up needed.

7 Best PC Motherboards for Prime Day Deals in 2026

Discover the best PC motherboards on Prime Day 2026, including options for AM4 and AM5 platforms, with detailed insights to help you choose the right build.

Wi-Wi is wireless time sync at 1 nanosecond

Wi-Wi, a Japanese-developed wireless synchronization protocol, demonstrates time accuracy down to 1 nanosecond, promising significant advances in broadcast and industrial tech.

Benchmarking Opus 5 On SlopCodeBench

Opus 5 audio codec was tested on SlopCodeBench, revealing performance benchmarks that could influence future codec development and adoption.