I Saw the Future of Robotics in a Virtual Duck That Chases My Cat

A Wired writer used OpenAI's GPT-6 Sol to train a simulated version of Microduck — a small robot from Pollen Robotics and Hugging Face — to waddle around a room and spot bengal cats, then set up a plan to dispense a treat for his pet bengal, Leona, possibly by bowing its head to release one from a 3D-printed storage chamber mounted on the robot's head. The physical Microduck hasn't shipped yet, but he expects the simulated behavior to carry over once it does.

Coding and training the virtual robot took only a few hours, work he said previously might have taken weeks. The speed traces to how well frontier models now handle 3D representations of physical space and convert them into code, agentic actions, and robot training data. Alongside GPT-6 Sol, the setup leaned on an agentic harness from Nous Research called Hermes and an Nvidia DGX Spark, a personal machine built around a powerful AI chip.

The model did more than write scripts. Using Nvidia's Isaac Sim, GPT-6 Sol fed instructions to the controller running the virtual robot, reviewed footage of its performance to judge what was working, and proposed adjustments to the training regimen — an iterative loop that got the Microduck walking without falling over. A failed side quest remains: after being shown a GIF of Michael Jackson moonwalking, hours of effort produced only what he called a jaunty backward jog.

Professional robotics labs are pursuing the same leverage. After OpenAI released Astra this past September, UC Berkeley professor Ken Goldberg and his students tested whether its 3D rendering ability could accelerate simulation work, the detailed and accurate environments needed so behaviors like a dancing humanoid or a grasping hand transfer faithfully to hardware. One lab member handed Astra an Nvidia tool called Newton, showed it a video of a real robot, and asked for a matching simulation.

Goldberg said the result stunned the group: within an hour, Astra had produced a complete 3D simulation with camera position, lighting, background, and materials all configured. He also credits Astra with automating much of the training drudgery, running experiments and inspecting the outcomes on its own.

The open question is whether these gains survive contact with hardware. The Microduck's cat-chasing and treat-dispensing still need validation on a physical unit that hasn't arrived, and sim-to-real fidelity remains the field's recurring bottleneck even as AI compresses setup time from weeks to hours.