OpenAI Folds Its Robotics Team Into Core Research as GPT-6 Astra Shows Robot Skills

This is the second time OpenAI has dissolved a robotics unit, but the reasoning is the opposite of last time. In 2020 the company shut the division because, co-founder Wojciech Zaremba said, it lacked training data and language models looked like a faster route to AGI. This time the company appears to think its general-purpose models, such as GPT-6 Astra, can control robots directly, so separate robotics models matter less. The aim is to speed up robotics work, not drop it.

The decision came fast. Fortune reported that OpenAI had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, and it stood empty.

What Astra Has Shown

The case for the change rests on recent tests of Astra. In an Nvidia study, an agent powered by Astra succeeded on 74.8 percent of simulated robotics tasks, against 51 percent for the best specialised models. Those tests used a simulated DROID robot, not physical hardware.

On real arms, a new benchmark called StationeryBench compared Astra with Ai2's MolmoAct2 on five desk-object tasks using dual-arm YAM robot arms. Over 200 trials, Astra fully completed 7 of its 100 tasks and MolmoAct2 completed none. Astra's average progress score was 46, against 12 for MolmoAct2. In a separate demonstration, Astra running on a humanoid robot moved around a kitchen and cleaned it up.

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, called Astra a major advance in spatial understanding. He said it reaches near-human accuracy on REMAP, a benchmark that has not yet been published. Artzi suspects OpenAI trained the model on large amounts of 3D data such as Blender scenes. OpenAI has not said how much robotics data went into its training.

The Limits of the Bet

General models still have weaknesses here. Specialised robotics models may keep advantages in cost, reliability, speed and energy use even if they generalise less well. Recent models from Anthropic and Google have also improved on robotics tasks, so OpenAI is not alone in moving robot control into frontier models.

OpenAI rebuilt its robotics team in 2024. Chief executive Sam Altman is reported to be aiming for a first public robot demonstration in 2027, and the company wants to build its own everyday robots over the long term. It also faces financial pressure to concentrate on its core AI business, where it has lost ground to Anthropic. That pressure was visible this week when OpenAI raised $15 billion, half its original target. The reorganisation means OpenAI's robotics progress now depends on how its flagship models improve, with no separate team working on robot-specific models.