DEV Community

Ai Game Changer Pro
Ai Game Changer Pro

Posted on Originally published at aigamechangerpro.blogspot.com

OpenAI Disbands Its Robotics Team — GPT-6 Astra Will Run the Robots Instead

OpenAI has broken up its standalone robotics team. Researchers now sit inside the main research group. Applied robotics staff moved to product. The company says it is not a retreat: its general-purpose model GPT-6 Astra scored 74.8% on an Nvidia robotics benchmark — against 51% for specialised models. The bet is that the model itself can now run robots.

Here is what happened, and why it matters.

OpenAI's robotics team is gone — as a team

OpenAI has disbanded its standalone robotics division. According to a report by Semafor, the researchers on the team have been folded into OpenAI's main research group, while the applied robotics staff — the people building and testing physical systems — have moved into the product organisation.

The team was only two years old. It was rebuilt in 2024 after OpenAI had shut down an earlier robotics effort in 2021, back when the company decided the hardware side was too slow and too expensive for a research lab.

That rebuild was serious. OpenAI hired from Meta, Tesla and the humanoid startup Figure, and in January 2025 it posted job listings for a humanoid robot programme. The plan was clear: build the robot, not just the brain.

The official line: this is not a retreat

OpenAI's response to the news is that nothing has been abandoned. A spokesperson said the company is "expanding our robotics efforts", and that the restructure puts robotics researchers closer to the frontier models they need.

The argument is straightforward. The old structure assumed that robotics needed its own team, its own stack, its own research agenda. OpenAI now believes the bottleneck is the model itself — perception, reasoning, planning — and that those problems are best solved inside the main research group, where the biggest models are built.

In other words, robotics is no longer a side project. It is a downstream application of the same model that writes code and answers questions.

The number behind the decision: 74.8%

The claim that supports this is a benchmark result. GPT-6 Astra, OpenAI's general-purpose model, scored 74.8% on an Nvidia robotics benchmark. Specialised robotics models scored 51%.

If those numbers hold up, the logic is hard to argue with. A general model that has never been trained specifically for robots beats the purpose-built ones by more than twenty points. Why maintain a separate robotics research team when the general model is already better at the task?

There is a caveat, and it is a big one. A benchmark is a test, not a robot. Scoring well on a simulated manipulation task is not the same as picking up a cup in a kitchen you have never seen, with a child running past. Benchmarks measure what is measurable. The messy part of robotics — the hardware, the edge cases, the safety — does not show up in a score.

What happens to the humanoid?

The first public humanoid demo is still targeted for 2027. That date has not moved, at least publicly. The restructure does not appear to have changed the hardware timeline.

What may have changed is who builds it. With the applied staff moved to product, the humanoid is now a product programme, not a research programme. That means schedules, budgets and shipping pressure — the things research teams are usually protected from.

It also means OpenAI will likely lean on partners for hardware. The company has invested in humanoid startups, and it has a history of letting others build the physical product while it supplies the intelligence. The January job postings suggest it still wants in-house hardware expertise, but the balance has shifted.

The broader pattern

This is not the first time a big AI lab has decided robotics is a model problem, not a hardware problem. Google DeepMind made the same bet with its Gemini Robotics work, arguing that a general model trained on video and language could control a robot arm with minimal task-specific training. Meta has moved in the same direction.

The pattern is consistent: the labs that win at models are deciding they do not need to win at motors.

That leaves the hardware to specialists — Figure, Tesla, Unitree — and the intelligence to the model labs. It is a division of labour that makes sense on paper. Whether it works in practice depends on something benchmarks cannot tell you: whether a model that has only ever seen the world through a camera can reliably move through it.

The bottom line

OpenAI has not quit robotics. It has decided that robotics is a model problem, and that the model team should own it. The 74.8% benchmark score is the justification. The 2027 humanoid demo is the test.

If the bet pays off, the next generation of robots will not be built by robotics teams. They will be built by product teams, running a general model. If it does not, OpenAI will have to rebuild the team it just disbanded — for the third time.

Related reads

Top comments (0)