I Saw the Future of Robotics in a Virtual Duck That Chases My Cat
A WIRED reporter successfully trained a virtual Microduck robot to chase cats using OpenAI's GPT-6 Sol, demonstrating how AI can dramatically accelerate robot development.
Using OpenAI's GPT-6 Sol, WIRED reporter Will Knight trained a virtual Microduck robot to identify and chase Bengal cats in a room. The virtual robot, developed by Pollen Robotics and Hugging Face, was taught to walk around and locate felines, with plans to eventually dispense treats to the reporter's pet cat, Leona.
The virtual training was completed in just a few hours, a process that previously might have taken weeks. This acceleration is possible because OpenAI's latest models excel at understanding and manipulating 3D representations of the physical world, translating them into useful code and robot training.
3D Simulation Breakthrough
When OpenAI released Astra in September, University of California, Berkeley professor Ken Goldberg and his students recognized its potential for robot simulations. These simulations require extreme detail and accuracy to ensure complex behaviors transfer faithfully to physical robots.
What blew us all away was that within an hour it had produced a full 3D simulation with all the parameters set—the position of the camera, the lighting, the background, all the materials.
Automated Robot Training
Astra can automate much of the drudgery involved in robot training by running experiments and examining results. Knight used GPT-6 Sol alongside Nous Research's Hermes agentic harness and an Nvidia DGX Spark computer to train the Microduck to walk without falling.
The AI used Nvidia's Isaac Sim to control the virtual robot, assess its performance through video analysis, and suggest training improvements. When Knight attempted to teach the robot to moonwalk using a Michael Jackson GIF, the AI produced only a "jaunty backward jog" after hours of effort.
New Frontiers in Robotics
Nvidia researcher Jim Fan compares how AI now controls robots to software agents functioning on a computer. Visual Language Action (VLA) models enable AI to manipulate physical actions similarly to accessing different software programs.
Fan, Goldberg, and others recently open-sourced Aspire, an agentic framework allowing robots to discover new skills. In one demonstration, GPT-6 Astra learned to use robot arms to solve a Rubik's cube through trial and error.
Practical Applications
Despite the vision system on Knight's robot duck not being reliable enough for real products, the ease of AI-powered robot training could soon impact manufacturing and e-commerce. Several industrial robotics and automation firms have shown interest in this approach.
The development may lead to "vibe robotics" breakthroughs as hobbyists and engineers discover how accessible programming and training robots has become. Knight hopes to perfect his robotic helper to ensure his cat Leona never lacks treats.
Related
