Wired's Will Knight uses GPT-6 Sol to train virtual Microduck

Wired's Will Knight trained a simulated Microduck, the $399 biped from Hugging Face's Pollen Robotics arm, to walk and chase his Bengal cat, Leona, with OpenAI's GPT-6 Sol managing the training loop. The reporting describes how 'an OpenAI model helped teach the Microduck to waddle and moonlight as a tiny cat butler.'

The rig paired GPT-6 Sol with Hermes, an agentic harness from Nous Research, running on an Nvidia DGX Spark desktop. The model wrote code, watched the duck's attempts, judged what was going wrong, and proposed revisions. Training ran across 4,096 parallel virtual ducks, which matches Pollen's own recipe for producing a usable gait in one to two hours. Two of the experts we track shared the piece.

The experiment never left the simulator. Knight had not received the physical robot, and reported 'its vision was not good enough for product deployment.' The treat-fetching cat butler, for now, is a plan.