OpenAI Disbands Robotics Team: GPT-6 Astra Completed 74.8% of Tasks vs. 51% for Narrow Models
OpenAI has disbanded its standalone robotics team, arguing that a general-purpose model, GPT-6 Astra, can control robots on its own. In an Nvidia test, it used camera images to control a virtual robot and completed 74.8% of object-handling tasks, while the best narrow robotics models managed 51%. But this was a simulation, and the figures have not been verified on physical robots. That makes data from real factory floors more valuable.
AI-processed from 3DNews AI; edited by Hamidun News
OpenAI has disbanded its standalone robotics team because it believes the same general-purpose model that powers chat and writes code, GPT-6 Astra, can control robots. The headline number comes from an Nvidia test. There, GPT-6 Astra received camera images, a description of the robot, and a command interface. It then controlled a virtual manipulator and completed 74.8% of tasks involving moving and handling objects. The best published result for specialized robotics models on the same leaderboard is 51%. An important caveat: all of this took place in simulation, not on real robots. For anyone building robots or investing in robotics startups, this signals that the market may shift from narrow models toward general-purpose ones.
What Exactly Changed at OpenAI
The company insists it is not abandoning the goal of building general-purpose robots. It simply sees no point in developing separate AI models for them. According to an OpenAI spokesperson, GPT-6 Astra unexpectedly performed well on certain robotics tasks.
Earlier this year, researchers expected this work to rely on the Sora video generator, which can model the physical world. Later it became clear that leading general-purpose models could be applied to robots directly. A separate team was no longer needed.
People were not laid off but reassigned to other groups. Researchers joined the core research division led by research chief Mark Chen. Specialists in applied robotics and hardware moved to the applied engineering division, which reports to OpenAI President Greg Brockman. Robotics went from a separate branch to one application of the shared model.
How GPT Controls a Robot Without Special Training
The scheme is easy to picture with an everyday analogy. The model is not trained on manipulator motion data. It is shown images from cameras, told what robot it is looking at, and given an interface through which it sends commands. It then acts like a person who has been briefed on an unfamiliar machine: looks at the scene, decides what to do, and issues a command.
In Nvidia's test, this approach completed 74.8% of object-manipulation tasks, compared with 51% for the best specialized models. The gap is notable. But the simulation has no friction, backlash, sensor noise, or communication delays, which a real robot must contend with. For that reason, these numbers cannot yet be transferred to physical devices.
What This Means for Robotics Startups
Many startups have bet on specialized models with "spatial intelligence." OpenAI's approach calls that bet into question. However, startups have something AI labs lack: their robots already operate in real conditions, such as factory floors, and they collect data on how things work in practice.
Russ Tedrake, co-founder and CEO of Walden Robotics, argues that the better the base models become, the more valuable site-specific data becomes. A strong model learns from that data faster, and whoever has collected it stands to benefit.
Large models also have limitations. A request is not processed instantly, and a cloud service requires a constant network connection. For a manipulator that must react within fractions of a second, or for a warehouse robot without reliable internet, this is a serious problem. Guanya Shi, associate professor at the Carnegie Mellon University Robotics Institute, therefore believes compact specialized models are still needed because they run directly on the robot's onboard computer.
Pete Florence, co-founder of Generalist, a startup building general-purpose models for robots, acknowledges that it is still unknown which approach will win in the long term.
Practical Takeaways
Developers and product teams should take three points from this news.
- A robot-control prototype built on a general-purpose model can now be assembled quickly, without separate training on robotics data.
- Simulation results are useful for comparing models but do not replace testing on real hardware.
- Competitive advantage increasingly lies in proprietary production data and in the ability to run models where fast internet is unavailable.
The market will likely go hybrid. A large model will plan and make decisions, while a small onboard model executes fast movements. OpenAI has dropped its standalone team, but the underlying pattern of "a general-purpose brain plus specialized skills" may well persist.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
What Exactly Changed at OpenAI?
The company insists it is not abandoning the goal of building general-purpose robots. It simply sees no point in developing separate AI models for them. According to an OpenAI spokesperson, GPT-6 Astra unexpectedly performed well on certain robotics tasks. Earlier this year, researchers expected this work to rely on the Sora video generator, which can model the physical world. Later it became clear that leading general-purpose models could be applied to robots directly. A separate team was no longer needed.
How GPT Controls a Robot Without Special Training?
The scheme is easy to picture with an everyday analogy. The model is not trained on manipulator motion data. It is shown images from cameras, told what robot it is looking at, and given an interface through which it sends commands. It then acts like a person who has been briefed on an unfamiliar machine: looks at the scene, decides what to do, and issues a command. In Nvidia's test, this approach completed 74.8% of object-manipulation tasks, compared with 51% for the best specialized models. The gap is notable.
What This Means for Robotics Startups?
Many startups have bet on specialized models with "spatial intelligence." OpenAI's approach calls that bet into question. However, startups have something AI labs lack: their robots already operate in real conditions, such as factory floors, and they collect data on how things work in practice. Russ Tedrake, co-founder and CEO of Walden Robotics, argues that the better the base models become, the more valuable site-specific data becomes.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.