# GPT-6 Sol trained a virtual robot duck in hours, and roboticists say "vibe robotics" is coming

> A WIRED writer taught a simulated Microduck to walk and spot his cat. At Berkeley, Astra built a full robot simulation from a video clip in about an hour.

- URL: https://devisdead.ai/news/gpt-6-sol-trained-a-virtual-robot-duck-in-hours-and-roboticists-say-vibe-robotic
- Published: 2026-11-11T20:22:36Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: robotics, openai, simulation, nvidia, agents
- Originally reported by WIRED (https://www.wired.com/story/i-trained-a-microduck-to-be-my-cats-butler/).

Will Knight spent a few hours teaching a virtual robot duck to find his cat, and came away thinking AI is about to make robot training much easier, he writes for [WIRED](https://www.wired.com/story/i-trained-a-microduck-to-be-my-cats-butler/).

## The duck

Using OpenAI's GPT-6 Sol, Knight trained a simulated Microduck (the little Hugging Face and Pollen Robotics bot) to stay upright as it wanders a room, and to recognize Bengal cats. He ran it with Nous Research's Hermes agent harness on an Nvidia DGX Spark. The model did more than write code: it drove Nvidia's Isaac Sim, watched video of the robot to judge how training was going, and suggested changes.

Knight says that took hours where it once would have taken him weeks. His goal is a real Microduck that bows to drop treats for his cat, Leona. His attempt to teach it a moonwalk from a Michael Jackson GIF produced only a jaunty backward jog.

## The lab version

UC Berkeley professor Ken Goldberg's lab gave OpenAI's Astra access to Nvidia's Newton tool and a video of a real robot, and asked for a simulation of the scene. Within about an hour it produced a complete 3D simulation with camera position, lighting, background and materials all set, Goldberg said. Astra can also run training experiments and review the results, taking over much of the tedious part.

Nvidia researcher Jim Fan compares it to software agents: a language model can call robot-control models, the vision-language-action kind, the way an agent calls apps. Fan and Goldberg were among the researchers who released Aspire as open source this week, a framework that lets robots discover and use new skills. In one demo, Astra worked out how to solve a Rubik's cube with two robot arms after a long stretch of trial and reasoning. At the Conference on Robot Learning in Austin, Fan showed Astra painting the Golden Gate Bridge with a robot arm.

## The catch

Reliable real-world robots still take serious engineering. Knight admits his duck's vision isn't good enough for a product. Still, Goldberg says industrial automation companies are interested, and Knight expects hobbyists to start doing for robots what vibe coding did for software.

**[Scoreboard](/scoreboard) update:** AI +1. AI models now set up and run robot training that used to take experts weeks.

Stay undead.