OpenAI moves robotics specialists to other teams

OpenAI is moving robotics researchers to other teams and expects general-purpose models, including GPT-6 Astra, to accelerate its work on robots. The researchers will join the core research team led by Chief Research Officer Mark Chen, while the applied robotics group and hardware team will move to Greg Brockman’s division. On simulated robotics tasks, an Astra-based agent scored 74.8%, compared with 51% for the best specialized models. In a test with real robotic arms, Astra completed 7 of 100 tasks.

OpenAI redistributes its robotics team across other divisions

Why OpenAI is changing its structure

The decision came so quickly that OpenAI had already booked a booth at the CoRL conference in Austin, but by the time the event began, the booth was empty. According to Fortune, the company believes that general-purpose models such as GPT-6 Astra will be able to control robots directly and reduce the need for separate robotics models.

In a study by Nvidia, an Astra-based agent completed 74.8% of tasks in simulation. The best result among specialized models was 51%. The tests used the simulated DROID robot, not a physical device.

A separate benchmark, StationeryBench, compared GPT-6 Astra with Ai2’s MolmoAct2. The models performed five tasks involving objects on a table, using YAM dual-arm robotic manipulators. Across 200 trials, Astra fully completed 7 of 100 tasks, while MolmoAct2 completed none. The average progress score was 46 for Astra and 12 for MolmoAct2.

In another test, Astra operated a humanoid robot that moved around a kitchen and cleaned it.

It is not yet known how much robotics data OpenAI and other companies used to train their models. General-purpose models from Anthropic and Google have also recently made progress on robotics tasks. Specialized models may still have advantages in cost, reliability, operating speed and energy use, even if they are less able to transfer skills to new situations.

Spatial understanding and the company’s plans

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, considers Astra a major step forward in spatial understanding. He says that on the unpublished REMAP benchmark, GPT-6 Astra achieves accuracy close to human performance. Artzi speculates that OpenAI may have trained the model on large volumes of 3D data, such as scenes from Blender.

OpenAI shut down its robotics division in 2020. Company co-founder Wojciech Zaremba explained that the team lacked training data and that large language models offered a faster path to artificial general intelligence. In 2024, the company reassembled the team.

OpenAI CEO Sam Altman reportedly hopes to hold the first public robot demo in 2027. In the long term, the company wants to build its own robots for everyday tasks. At the same time, financial pressure is pushing OpenAI to focus on its core AI business, where it has fallen significantly behind Anthropic.

Dataist’s take: The restructuring reflects OpenAI’s bet on general-purpose models as the basis for robot control, but the reported results do not prove they perform better in real-world conditions: some tests took place only in simulation, and specialized models may be cheaper, more reliable and more energy-efficient. The company is still pursuing plans to build robots and hold a public demo, even as pressure mounts to focus on its core business.