🔍 Read the full analysis: Do Robots Need Specialized AI? OpenAI's Robotics Team Split Sparks Debate on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI confirmed that it has moved its robotics researchers and applied robotics teams into other parts of the company, saying general-purpose models can now be applied directly to robot control. A simulation study found GPT-6 Astra outperformed specialist policies on task success, while Runway’s Praxis-1 acted much faster in reported tests. The results leave open whether robots will rely on one kind of model or combine both.
OpenAI has broken up its stand-alone robotics team and moved its researchers and applied robotics staff into other company divisions, according to Fortune, which said OpenAI confirmed the change. The company told Fortune that general-purpose models such as GPT-6 Astra can now be applied directly to robotics, raising questions about the role of specialized AI in robot control.
Fortune reported on November 13 that OpenAI’s robotics researchers now sit within its main research organization under Chief Research Officer Mark Chen. Applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. Fortune said an OpenAI spokesperson described the reorganization as a commitment to robotics, rather than a retreat. OpenAI has not published a separate statement about the move, according to the source material.
The stated rationale is that frontier models have developed capabilities that OpenAI believes can be used for robot control without relying on a stand-alone robotics group. A simulation study by researchers at NVIDIA’s Seattle Robotics Lab reported that an agent built around GPT-6 Astra achieved a 74.8% success rate on the RoboLab-120 benchmark in an agentic setup. The study compared that result with 51.0% for the best trained robot policy on the benchmark’s leaderboard. These are results from a simulation, not a demonstration of performance in everyday physical settings.
The same study reported a significant speed limitation: in its direct setup, GPT-6 Astra calls took an average of 18.56 seconds, with the simulation pausing while the model processed its response. Runway, meanwhile, released Praxis-1, a 27-billion-parameter world action model under the Apache 2.0 license. It scored 46.2% on RoboLab-120 and took roughly 0.55 to 0.6 seconds per call for 1.6 seconds of motion, according to the source material. The results come from different reported evaluations and should not be treated as a single controlled comparison.
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 and Skill Pull Apart
The reported findings show different strengths in the tested setups, rather than a clear overall winner. General-purpose models showed higher task success in one simulation setup, while dedicated robot models were designed to respond at a pace more compatible with continuous physical action. In the cited study, GPT-6 Astra calls took an average of 18.56 seconds in a direct setup, with the simulation paused during processing.
The results are relevant to decisions about how robot systems are built. General models may offer flexible reasoning across tasks, while specialized policies may provide faster control. Runway reported that, in a test on one of its robots, GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. That comparison is Runway’s own test, not an independently reported head-to-head study, and does not establish performance across other robots or tasks.
The reorganization reflects OpenAI’s stated view that general models can be applied to robotics. Runway’s release presents a separate approach based on a model designed for action. The reported results do not establish which approach will be used more widely or whether they will be combined.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Bose QuietComfort Wireless Bluetooth Headphones – Active Over‑Ear Noise Cancelling with Mic, USB‑C Charging, Deep Bass, up to 24 h Playtime,
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
OpenAI’s Shifting Robotics Strategy
The reorganization follows earlier changes in OpenAI’s robotics effort. The company wound down a stand-alone team in 2020, then rebuilt one in 2024 and partnered with humanoid robot maker Figure. Figure said in February 2025 that it would develop its own robot AI, ending that partnership. Fortune reported that OpenAI’s earlier research plans had considered building on video models such as Sora; the spokesperson’s more recent explanation was that general-purpose frontier models had shown direct potential for robotics.
OpenAI has also said it is developing a humanoid robot. CEO Sam Altman has targeted a first public demonstration in 2027, according to the source material. Fortune reported that OpenAI’s booth at the CoRL robotics conference in Austin remained on the floor plan but was unoccupied when the publication visited. The source also notes broader cost and product-focus pressures at OpenAI, but the company’s stated reason for the team change was its view of general models’ robotics capabilities.
“Complementary.”
— Runway, describing its view of the two approaches
Real-World Performance Still Open
The benchmark figures cited so far do not establish how either approach performs across physical robots, varied environments or longer tasks. NVIDIA’s reported GPT-6 Astra result came from a simulation involving 120 tabletop tasks for a single robot arm, and the study’s slower direct setup highlights a latency issue. It is not clear how the reported success advantage translates to real-time use or whether additional systems can reduce the delay.
The Praxis-1 and GPT-6 Astra figures also come from distinct evaluations. Runway’s speed and cost comparison involved one of its robots and three tasks, according to the source material; the details needed to judge how broadly that result applies are not provided here. OpenAI has not publicly detailed the reorganization in its own announcement, and the source does not establish what research projects or staffing changed beyond the teams’ new reporting lines.
It also remains uncertain whether OpenAI will use general models alone, retain specialized components within its broader organizations, or combine both. The source material describes the possibility of cooperation between general and specialist models but does not report a final system design.
Evidence to Watch in 2027
Further disclosures from OpenAI about its robotics organization and physical-robot tests could provide additional evidence about how its models perform. OpenAI’s stated target of a public humanoid-robot demonstration in 2027 could indicate how the company plans to apply its models, though the source material does not confirm a date or what the demonstration will show.
Further testing could assess whether general models can reduce response times without losing task performance, and whether specialized models can handle a wider range of instructions. The reported studies show different strengths for general models and dedicated robot models, but do not establish whether one can replace the other.
Key Questions
What did OpenAI change?
Fortune reported that OpenAI moved its robotics researchers into its main research organization under Mark Chen and its applied robotics and hardware teams into the Applied organization under Greg Brockman. OpenAI confirmed the change when Fortune asked.
Why did OpenAI reorganize its robotics teams?
According to Fortune, an OpenAI spokesperson said general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. OpenAI framed the move as part of its continued commitment to robotics.
Did GPT-6 Astra outperform specialist robot models?
In an NVIDIA researchers’ simulation study, an agentic GPT-6 Astra setup recorded a 74.8% success rate on RoboLab-120, compared with 51.0% for the best trained policy on the leaderboard cited in the source. That is a simulation result, not proof of superiority in physical robot use.
What is Praxis-1?
Praxis-1 is Runway’s 27-billion-parameter world action model, released with weights and code under the Apache 2.0 license, according to the source material. It scored 46.2% on RoboLab-120 and had much shorter reported call times than GPT-6 Astra in the cited comparisons.
Do robots need specialized AI?
The available evidence does not settle the question. General-purpose models have shown strong task success in simulation, while specialized models have shown faster response times in reported tests. The approaches may be used together, but no final design is established in the source material.
Source: ThorstenMeyerAI.com
