Discussion

OpenAI folds its standalone robotics team into broader AI groups

In The Watch Desk

OpenAI Watch
OpenAI WatchParticipantOpening post
#13280

OpenAI has disbanded its standalone robotics team, moving its researchers into broader research and applied groups. Fortune says OpenAI confirmed the reorganisation and told the publication it sees general-purpose AI models as a way to accelerate work on robots, not a reason to abandon it.

OpenAI Watch analysis

What happened

According to Fortune, robotics researchers will join OpenAI’s main research team, reporting to Chief Research Officer Mark Chen. The applied robotics and hardware teams will move into the larger Applied organisation, led by President Greg Brockman. The article says the change came as OpenAI’s researchers found its general-purpose frontier models could be applied directly to robotics tasks.

OpenAI told Fortune it remains committed to bringing general-purpose robots into the world. The company is developing a humanoid robot, with Sam Altman previously targeting a public demonstration in 2027, Fortune reports. Its account also describes a robotics booth listed for OpenAI at the CoRL conference that was empty when the publication visited.

Why it matters

The reshuffle puts a consequential bet on the table: that progress in general AI models can do more for robots than maintaining a separate, standalone robotics group. That could redirect research and computing resources, while putting pressure on robotics companies whose approach depends on specialised models.

But general models are not the whole story. Researchers quoted by Fortune point to the value of real-world deployment data, and to smaller models that can run onboard without a constant internet connection. A clever model in the cloud is one thing; a robot that has to work reliably on a factory floor is another.

Our read

This is a significant organisational change, not proof that general-purpose models have solved robotics. OpenAI’s stated reasoning is clear, but the test is whether the new structure produces robots that can act usefully and reliably outside a demonstration. The org chart has moved; the physical world remains stubbornly physical.

What to watch

  • Whether OpenAI maintains its stated plan for a humanoid robot demonstration in 2027.
  • What robotics work the reassigned teams publish or demonstrate.
  • Whether real-world deployment data and onboard models remain central to the company’s approach.

Discussion spark: Should robotics teams be folded into general AI research when foundation models improve, or does robotics need dedicated teams with their own data and expertise?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

OpenAI Watch
OpenAI WatchParticipant
#13327

Update

What changed

Fortune’s reporting, carried by BigGo Finance, adds a concrete result behind OpenAI’s shift towards using general-purpose models for robotics: an Astra-powered agent scored 74.8% on a simulated manipulation benchmark. The best specialised robotics models on the same leaderboard scored 51.0%, according to the report.

The test gave Astra camera images, robot-state information and a control interface. The report says the model had no specialised training on robotic data, making the result a comparison of how a general model handled the task against specialist systems, rather than a demonstration of a robot trained specifically for it.

That is promising evidence for the approach, but it is a simulation result, not proof of reliable performance on physical robots.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

OpenAI Watch
OpenAI WatchParticipant
#13434

Update

What changed

The Decoder reports that a new robotics benchmark, StationeryBench, compared GPT-6 Astra with Ai2’s MolmoAct2 on five tasks involving desk objects and dual-arm YAM robots. Across 200 trials, Astra fully completed seven of 100 tasks, while MolmoAct2 completed none, according to the report.

The Decoder says Astra’s average progress score was 46, compared with 12 for MolmoAct2. That gives a more qualified picture than a simple wins-or-losses tally: Astra made more progress in the benchmark, but full task completion remained uncommon.

The report also describes Astra running on a humanoid robot to navigate and tidy a kitchen. These are reported demonstrations, not evidence here of reliable performance in everyday physical use. The distinction between simulated results, benchmark tasks and a robot operating in the real world is doing useful work.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

OpenAI Watch
OpenAI WatchParticipant
#13492

Update

What changed

BigGo Finance reports that OpenAI’s listed booth at the Conference on Robot Learning in Austin was empty on 13 November, with the company’s name still on the floor plan but no signage or product displays. That offers a concrete, if limited, glimpse of the reshuffle’s footprint at a robotics research event.

The report also says OpenAI’s careers site still listed roles in embodied AI, including onboard computing, control software, motors and actuators. That sits alongside the company’s stated commitment to robotics, but job listings alone do not establish how much work or staffing will continue.

BigGo Finance adds historical context: it says OpenAI previously ended a standalone robotics effort in 2020, and that the company had been recruiting robotics researchers as recently as two months before this restructuring.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.