OpenAI reshapes robotics around its general AI models
OpenAI
The company behind ChatGPT.
embodied AI
AI that controls devices in the physical world.
simulation
A computer test that imitates a real task.
What happened
According to Fortune’s report, OpenAI, the company behind ChatGPT, has disbanded its standalone robotics team. It moved the researchers into other parts of the company. Robotics research now reports to the main research group led by Chief Research Officer Mark Chen. Applied robotics and hardware staff moved into the larger Applied organization, led by President Greg Brockman. An OpenAI spokesperson said robotics remains a long-term priority. The company described the change as a way to speed progress, not abandon robots.
The background
OpenAI had a robotics team in its early years. It wound that team down in 2020. Cofounder Wojciech Zaremba later pointed to limited robot data and a more efficient path through other AI work. OpenAI rebuilt a standalone robotics team in 2024. It also announced a partnership with Figure, a humanoid-robot startup. Figure said in February 2025 that it would develop its own robot AI. OpenAI has said it is now developing its own humanoid robot. CEO Sam Altman has targeted a first public demonstration in 2027.
Why the change matters
The reorganization signals a strategic bet. OpenAI appears to think robotics may no longer need a separate AI brain. General-purpose models already work with language, code, and images. The company says frontier models can also help control physical devices. This area is often called embodied AI. If the bet works, one model family could support many kinds of work, including robot control. Researchers could focus on improving a shared model, while other teams handle hardware and applications.
That idea has an important limit. A robot works in a changing physical world. Objects move. Lighting changes. People enter the space. A system must notice those changes and choose safe actions. Success with text or images does not automatically prove reliable control of a machine.
What is confirmed
OpenAI confirmed the reorganization when Fortune asked about it. Fortune also visited OpenAI’s registered booth at CoRL, a major robotics conference in Austin. The booth was empty and had no OpenAI logo. A later update clarified that former members of the team were still at the conference.
The article also described a Nvidia study. Researchers gave GPT-6 Astra camera images, robot information, and a control interface. An Astra-powered agent scored 74.8% on a simulated manipulation benchmark. The best specialized robot models reported on that leaderboard scored 51.0%. The test happened in a computer simulation. It was not a test of a humanoid robot working in a factory.
The story drew 121 points and 54 comments on Hacker News, a technology discussion site. Those numbers show community attention. They do not prove that the report is true.
What remains unknown
The report does not show how often a general model can control a real robot safely. It also does not give a detailed schedule for the new teams. Large models can be slow and may need a constant cloud connection. Robotics companies with deployed machines have another advantage: real work creates useful data. Experts also say smaller, specialized models may still be needed on the robot itself.
What to watch next
The clearest test will be a real demonstration. OpenAI has said it wants to show a humanoid robot in 2027. Watch whether it repeats tasks safely, handles unexpected changes, and works without a constant network connection. The key question is not whether OpenAI changed an org chart. It is whether general models can produce dependable results in the physical world.
OpenAI is changing how it works on robots
📰 Full story: OpenAI reshapes robotics around its general AI models
OpenAI moved its robotics researchers into larger research and product groups.
general-purpose model
An AI model built for many kinds of work.
simulation
A computer test that imitates a real task.
Hacker News
A website where people discuss technology.
💡 The gist
- OpenAI, the company behind ChatGPT, disbanded its standalone robotics team.
- It wants general AI models to help control robots.
- The idea is promising, but real-world success remains unproven.
What changed?
Fortune reported that OpenAI moved robotics researchers into its main research group. Applied robotics and hardware staff moved into its Applied organization. The company says this will speed progress. It also says robotics remains a long-term goal.
OpenAI once had a robotics team. It closed that team in 2020. It rebuilt a standalone team in 2024. It later lost a partnership with Figure in 2025. OpenAI now says it is building its own humanoid robot. It has mentioned a public demonstration in 2027.
Why change the team?
A general-purpose model can work with language, code, and images. OpenAI now says these models can also help with physical machines. This could reduce the need for a separate robot brain. It could also let one model support many different jobs.
Robots still face difficult situations. A bottle can move. Light can change. A person can enter the room. The AI must notice each change. It must then choose a safe action. This is harder than answering a question on a screen.
What do the numbers show?
A Nvidia study tested GPT-6 Astra with a simulated robot. The agent scored 74.8%. Specialized robot models scored 51.0%. The test happened inside a computer simulation. It did not prove that a humanoid robot works reliably in a factory.
Large models may also be slow. They may need a constant internet connection. Smaller models could still help robots work onboard.
The story reached 121 points and 54 comments on Hacker News, a technology discussion site. Those numbers show attention. They do not prove the report is true.
What should readers watch?
Watch for a real robot demonstration. Check whether it repeats tasks safely. Check whether it handles changes around people. Also check whether it works without constant cloud access.
💬 Can one big AI run a robot?
The discussion asked both whether one general AI can control a robot and why OpenAI changed its robotics organization.
- A big AI may eventually beat special-purpose systems by using more data and computing. But robots do not always have cheap, plentiful computing or enough speed for every decision.
- A robot is more likely to use layers: a big model chooses the goal, small fast systems make movements, and hardware-level controllers move the motors.
- One user self-reported that a game-playing model chose its next action in 36ms. Another commenter said motor control may need about 1kHz, or one cycle per millisecond. These are user reports or judgments, not independent measurements. The robot should also have a local safety stop if the network disappears.
- Older tools such as PID, EKF, and homography methods can still work alongside general models. A new general model does not automatically replace every specialized controller.
- The team breakup is not proof that robotics has been solved. Commenters said researchers moved to other teams and that a registered CoRL booth was empty; some saw a priority or cost shift, while others guessed at IPO-related savings.
initial digest at 54 comments (revision 1). We fetched 54 comments and sampled 54 across the thread. These are HN users’ reports, not independently verified facts.
How will OpenAI teach robots?
📰 Full story: OpenAI reshapes robotics around its general AI models
OpenAI changed where its robot workers sit.
OpenAI
The company that makes ChatGPT.
general AI
AI that can help with many jobs.
Hacker News
A website where people talk about technology.
OpenAI, the company that makes ChatGPT, changed its robot team. The team was split among bigger teams. OpenAI says it still wants to make robots. It says a general AI might help robots.
A robot sees things. A robot moves things. It must choose the next safe move. OpenAI thinks one AI can help with many jobs.
A computer test used GPT-6 Astra, an AI model. It got 74.8 percent. Another model got 51.0 percent. 74.8 is bigger than 51.0. But the test was inside a computer. It was not a real factory test.
Big AIs can be slow. They may need the internet. Smaller AIs may still help robots.
OpenAI says it wants to show a human-shaped robot in 2027. We should watch that real test.
The story was discussed on Hacker News, a technology website. It got 121 points and 54 comments. Those numbers show attention. They do not show truth.
💬 A robot needs a big brain and fast little helpers
A big AI can think about the goal, but small fast helpers should move the motors and stop danger.
- The big AI can decide what the robot should do. Small helpers can move its hands and feet, and a safety device inside the robot can stop it if the connection breaks.
- One user self-reported that a game AI chose a move in about 36ms. Another said motors might need about 1ms. These are reports or opinions from commenters, not checked measurements.
- Old special tools such as PID can still help. They can work together with the big AI.
- Ending the robotics team does not prove that robots are solved. The company may simply have moved people to other teams and changed its priorities or spending.
initial digest at 54 comments (revision 1). We fetched 54 comments and sampled 54 across the thread. These are HN users’ reports, not independently verified facts.
💬 Did general-purpose models make robotics research unnecessary?
HN commenters split between seeing OpenAI’s robotics-team breakup as proof that general models have won, and seeing it as a change in priorities or business strategy.
initial digest at 54 comments (revision 1). We fetched 54 comments and sampled 54 across the thread. These are HN users’ reports, not independently verified facts.