OpenAI folds robotics researchers into research and Applied teams

OpenAI folds robotics researchers into research and Applied teams

OpenAI has reorganized its robotics team, according to The Decoder. The outlet first reported it as OpenAI "apparently disbanding its robotics team again", then updated the article after OpenAI clarified that the team was reorganized to work more closely with research and product, and that robotics work continues.

The split works like this. Robotics researchers are moving to the main research team under Chief Research Officer Mark Chen. The applied robotics and hardware teams join Greg Brockman's Applied organization. The Decoder describes the move as a way to accelerate robotics by folding researchers into other teams, not abandon it. Fortune reports that the decision came on such short notice that OpenAI had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, and the booth then sat empty.

The Decoder writes that OpenAI apparently believes general-purpose models like GPT-6 Astra can control robots directly, which would reduce the need for specialized robotics models. That is the outlet's inference, not an OpenAI statement. It points to several test results.

In an Nvidia study, an Astra-powered agent reached a 74.8 percent success rate on simulated robotics tasks, against 51 percent for the best specialized models. The tests used a simulated DROID robot, not a physical one.

A new benchmark called StationeryBench 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, while MolmoAct2 completed none. Astra's average progress score was 46, compared with 12 for MolmoAct2. In another test, Astra running on a humanoid robot navigated and cleaned up a kitchen.

The Decoder adds caveats. It remains unclear how much robotics data OpenAI and others included in training. Large general-purpose models from Anthropic and Google have also made recent progress on robotics tasks. Specialized robotics models may still win on cost, reliability, speed and energy consumption, even if they generalize less well.

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, also 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 On the still-unpublished REMAP benchmark, GPT-6 Astra reaches near-human accuracy, according to Artzi.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 OpenAI may have trained the model on large amounts of 3D data such as Blender scenes.

The history is mixed. OpenAI shut down its robotics division in 2020 because, according to co-founder Wojciech Zaremba, the company lacked the training data it needed and large language models were a faster path to AGI. It 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. The Decoder also notes that the company faces heavy financial pressure to focus on its core AI business, where it has lost significant ground to Anthropic.

Key facts

  • Robotics researchers move to the main research team under Chief Research Officer Mark Chen; applied robotics and hardware teams join Greg Brockman's Applied organization.
  • OpenAI told The Decoder the reorganization is meant to work more closely with research and product, and that robotics work continues. The outlet corrected its original 'disbanding' framing.
  • In an Nvidia study on a simulated DROID robot, an Astra-powered agent scored 74.8 percent on robotics tasks versus 51 percent for the best specialized models.
  • In StationeryBench, GPT-6 Astra fully completed 7 out of 100 tasks and MolmoAct2 completed none; average progress scores were 46 versus 12.
  • OpenAI shut its robotics division in 2020 and rebuilt it in 2024; Sam Altman is reportedly targeting a first public robot demo for 2027.

Why it matters

OpenAI closed its robotics division once before, in 2020, and rebuilt the team in 2024. This reorganization points to a different approach: instead of a separate robotics group building specialized models, researchers join the main research team and the hardware side joins Applied. The Decoder says OpenAI apparently believes general-purpose models like GPT-6 Astra can control robots directly. If that holds, specialized robotics models matter less. Test results cited in the article, including a 74.8 percent versus 51 percent gap in a simulation study, are the evidence offered for that idea.

Who it affects

OpenAI's robotics researchers and its applied robotics and hardware staff are the ones being moved, to the teams under Mark Chen and Greg Brockman respectively. Companies building specialized robotics models, such as Ai2 with MolmoAct2, are the comparison point in the benchmarks. Anthropic and Google also matter here: their general-purpose models have made recent progress on robotics tasks, the article says.

How to use it

There is nothing to adopt here. This is a report about an internal reorganization, not a product release. Readers following robotics AI can use the article as a pointer to the evaluations it names: the Nvidia simulation study, StationeryBench, and the still-unpublished REMAP benchmark that Yoav Artzi cites.

How solid is it

The reorganization itself is reported by The Decoder, which corrected its first framing after OpenAI said robotics work continues. The booth detail comes from Fortune. The claim that OpenAI thinks general-purpose models can replace specialized ones is The Decoder's inference, hedged with 'apparently', not an OpenAI statement. The Nvidia result is from a simulated DROID robot, not a physical one. The StationeryBench figures are given as stated: 200 trials, but 'out of 100 tasks', with no scale for the progress score. The REMAP benchmark is unpublished, and Artzi's guess about Blender-scene training data is his suspicion, not a confirmed fact. Altman's 2027 demo target is reported, not confirmed.

Risks and caveats

The article itself lists the caveats. It is unclear how much robotics data OpenAI and others used in training. Specialized robotics models may keep advantages in cost, reliability, speed and energy use, even if they generalize less well. In StationeryBench Astra fully completed only 7 out of 100 tasks, so strong relative results sit next to low absolute completion. The Decoder also notes OpenAI faces heavy financial pressure to focus on its core AI business, where it has lost significant ground to Anthropic, which could shape how much robotics work gets done.

“Rather than abandoning robotics, the company wants to accelerate it by folding researchers into other teams.”

— The Decoder