Agentic AI
AI-Powered Robots Learn Faster Than Ever, From Duck Butlers to Rubik's Cube Solvers
Advanced language models are dramatically accelerating robot training, letting researchers and hobbyists build sophisticated automated behaviors in hours instead of weeks.
WIRED — AI
The barrier to building intelligent robots is collapsing. What once took weeks of specialized knowledge can now happen in an afternoon, thanks to generative AI models that understand the physical world well enough to train robots in simulation.
The transformation is being driven by a combination of advances: large language models capable of writing and debugging code, visual AI systems that can render 3D environments, and frameworks that let these systems act autonomously to test and refine robotic behavior. The result is a democratization of robotics that could reshape manufacturing and automation.
From Hours to Weeks: Training Gets Dramatically Faster#
A Wired reporter recently trained a virtual version of Microduck—a small robot made by Pollen Robotics and Hugging Face—to autonomously identify and approach cats using OpenAI's GPT-6 Sol and an agentic framework called Hermes. What's striking: the entire project, from initial concept to a working behavior, took just a few hours. The robot learned to walk without falling over, and the developer plans to adapt the behavior for a physical unit that will dispense treats to cats by bowing its head.
This acceleration stems from AI models' newfound facility with 3D simulation. When OpenAI released Astra in September, researchers at UC Berkeley's Ken Goldberg lab tested whether it could generate detailed simulations from video footage. Using Nvidia's Newton physics engine, Astra consumed a video of a real robot and produced a full 3D simulation complete with camera position, lighting, materials, and environmental details—in under an hour. Crucially, the model then automated much of the trial-and-error process, running experiments and analyzing results to guide training.
The Intelligence Behind Physical Control#
Nvidia researcher Jim Fan notes that the way modern AI controls robots mirrors how software agents function: large language models can access visual language action (VLA) models trained to control physical movement much as they would call software tools on a desktop. This week, Fan and collaborators including Goldberg open-sourced a framework called Aspire, which allows robots to discover and apply new capabilities iteratively.
The demonstrations are striking. At the Conference on Robot Learning in Austin, Fan showed Astra operating a robot arm that painted the Golden Gate Bridge from scratch using available paint and brushes. In another test, GPT-6 Astra manipulated a Rubik's cube using a pair of robot arms—failing initially, reasoning through the problem over time, and eventually solving it by learning intermediate skills.

