Wired writer trains virtual Microduck robot to spot cats using OpenAI's GPT-6 Sol
A Wired writer says he used OpenAI's GPT-6 Sol model to train a simulated version of the Microduck robot, made by Pollen Robotics and Hugging Face, to walk around a room and identify bengal cats. In a newsletter edition, Will Knight said the coding and training took just a few hours, work that he said might previously have taken weeks.
Knight is still waiting for a physical Microduck. He said behavior learned in simulation is relatively simple and should carry over to the real robot. He plans to have it dispense a treat for his pet bengal, Leona, possibly by bowing its head to release one from a 3D-printed storage chamber mounted on it.
How the training worked
Knight paired GPT-6 Sol with Hermes, an agentic harness from Nous Research, and an Nvidia DGX Spark, a personal computer with a powerful AI chip. According to Knight, the model went beyond writing code: it used Nvidia's Isaac Sim software to send instructions to a model controlling the virtual robot, watched videos of the results to judge progress and suggested changes to the training approach.
The process taught the Microduck to walk without falling over. An attempt to teach it to moonwalk, using a GIF of Michael Jackson, ended after hours of effort in what Knight called a "very jaunty backward jog."
Knight cautioned that making robots work reliably remains very difficult and requires significant engineering expertise. The vision system on his robot is not good enough for a real product, he wrote.
Researchers see wider use
Knight cited work by Ken Goldberg, a professor at the University of California, Berkeley, whose students tried OpenAI's Astra model after its release in September. A lab member connected Astra to Newton, an Nvidia tool, supplied a clip of a real robot at work and asked the model to build a simulation of that scene.
“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,”
Goldberg also said Astra can automate much of the tedious work of simulation training by running experiments and examining results. He said several industrial robotics and automation firms have shown interest in the approach, according to Knight.
Jim Fan, an Nvidia researcher who has worked with Goldberg, said AI models can now control robots in a way that mirrors software agents. Large language models can call on vision-language-action models, which are trained to control physical actions, much as they use different programs on a computer.
This week Fan, Goldberg and others released an open-source agentic framework called Aspire, which lets a robot discover and use new skills. In one example, GPT-6 Astra worked out how to use two robot arms to solve a Rubik's cube. Fan also gave a keynote at the Conference on Robot Learning in Austin, Texas, showing Astra directing a robot arm to paint the Golden Gate Bridge.
