OpenAI Disbands Robotics Team Again — This Time Because GPT-6 Astra Might Not Need It

OpenAI has reassigned its robotics researchers to other teams, betting that general-purpose models like GPT-6 Astra can control robots directly and make specialized robotics AI unnecessary. The company made the decision on such short notice that it had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, which then sat empty, according to Fortune.

This is not OpenAI's first robotics breakup. But the reasoning has changed since 2020, and the stakes for the broader robotics industry have changed with it.

What exactly happened to the team?

OpenAI is not abandoning robotics, Fortune reports. Instead, the company wants to accelerate it by folding researchers into other parts of the organization.

The split works like this:

• Robotics researchers move to the main research team under Chief Research Officer Mark Chen.
• The applied robotics and hardware teams join Greg Brockman's Applied organization.

The reassignment arrived abruptly. OpenAI had committed to a presence at CoRL in Austin, one of the field's main academic gatherings, and the booth stood empty after the reshuffle.

The empty booth tells the story. A company that rebuilt its robotics effort in 2024 pulled its researchers out mid-cycle because it no longer believes a separate robotics team serves the goal.

Why does OpenAI think specialized robotics AI is obsolete?

The core argument is capability. OpenAI apparently believes that general-purpose models like GPT-6 Astra can control robots directly, reducing the need for specialized robotics models built only for machines.

Early numbers back that belief, at least partially.

In an Nvidia study, an Astra-powered agent achieved a 74.8 percent success rate on simulated robotics tasks. The best specialized models managed 51 percent. The tests used a simulated DROID robot, not a physical one — an important caveat for anyone extrapolating to real-world deployment.

A new robotics benchmark called StationeryBench adds more evidence. It compared GPT-6 Astra with Ai2's MolmoAct2 on five tasks involving desk objects, run on dual-arm YAM robotic arms. Across 200 trials, Astra fully completed 7 out of 100 tasks. MolmoAct2 did not complete a single one. Astra's average progress score was 46, against 12 for MolmoAct2.

In another test, Astra running on a humanoid robot navigated and cleaned up a kitchen.

The pattern matters beyond OpenAI. Other large general-purpose models from Anthropic and Google have also made recent progress on robotics tasks, suggesting the trend is structural rather than specific to one company's model.

How solid is the evidence?

The picture has gaps. It remains unclear how much robotics data OpenAI and others incorporated during training — a critical question, because benchmark wins mean less if the models were extensively trained on similar tasks.

Specialized robotics models may still hold advantages in several dimensions:

• Cost
• Reliability
• Speed
• Energy consumption

These matter enormously for commercial deployment, where a smaller, cheaper, more predictable model can beat a general one that scores higher on benchmarks but costs more to run. Specialized models may not generalize as well, but generalization is not the only metric that counts.

What do outside researchers say?

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, sees Astra as a major leap in spatial understanding. On the still-unpublished REMAP benchmark, GPT-6 Astra reaches near-human accuracy, according to Artzi.

He suspects that OpenAI may have trained the model on large amounts of 3D training data, such as Blender scenes. That suspicion, if confirmed, would connect back to the open question about training data — near-human spatial understanding looks different if it comes partly from heavy exposure to synthetic 3D environments.

OpenAI and robots: what is the history?

OpenAI shut down its robotics division in 2020. Co-founder Wojciech Zaremba said at the time that the company lacked the training data it needed and that large language models offered a faster path to AGI.

The company rebuilt the team in 2024. CEO Sam Altman is reportedly targeting a first public robot demo for 2027, and OpenAI wants to build its own everyday robots in the long term.

Now the company is dissolving the standalone team again — while keeping the goal. The bet is that the path to robots runs through the same general models that power its core products, not through a parallel research track.

There is also financial pressure in the background. OpenAI faces heavy pressure to focus on its core AI business, where it has lost significant ground to Anthropic. A separate robotics division is a luxury; general-purpose models that also happen to control robots are an extension of the main product line.

Why does this matter for the robotics industry?

If general-purpose models can match or beat specialized robotics AI, the economics of the entire robotics software stack shift. Startups building robot-specific models would compete against frontier models that improve every year for reasons unrelated to robotics. Embodied AI could become a feature of general models rather than a separate discipline.

The evidence is early — simulated robots, small task counts, unpublished benchmarks — but the direction is consistent across OpenAI, Nvidia's testing, and independent researchers like Artzi.

The next checkpoint is concrete: whether OpenAI's reported 2027 public robot demo arrives, and whether it runs on the same general models that just absorbed its robotics team.