OpenAI Models Cut Robot Simulation Setup to 1 Hour as Researchers Push New Agentic Framework
Updated
Updated · WIRED · Nov 11
OpenAI Models Cut Robot Simulation Setup to 1 Hour as Researchers Push New Agentic Framework
3 articles · Updated · WIRED · Nov 11
Summary
OpenAI’s latest models are sharply speeding robot training in simulation, with one Berkeley lab test generating a full 3D scene from a robot video clip in about an hour.
That gain comes from stronger 3D understanding and agentic control: the models can build simulations, write code, run experiments, watch robot performance, and suggest training tweaks inside Nvidia’s Isaac Sim and Newton tools.
In a hands-on test, GPT-6 Sol helped train a virtual Microduck robot in a few hours to walk around and identify Bengal cats—work the author said previously could have taken weeks.
Researchers are already extending the approach: Nvidia’s Jim Fan, Berkeley’s Ken Goldberg and others this week open sourced Aspire, a framework for robots to discover and use new skills, including solving a Rubik’s cube.
At CoRL in Austin, demos included Astra directing a robot arm to paint the Golden Gate Bridge, though researchers said reliable real-world deployment still needs heavy engineering even as industrial interest grows.