WIRED writer trains a virtual robot duck with AI—but its vision falls short

A few hours of AI-assisted work produced a virtual robot duck that could walk without falling and identify Bengal cats, WIRED’s Will Knight reported on November 11. His experiment used OpenAI’s GPT-6 Sol to help manage training inside a simulator—a step beyond generating code, though not yet a demonstration on a physical robot.

Knight’s subject was a simulated Microduck, a small robot from Pollen Robotics and Hugging Face. He was still waiting for a physical unit when he published the experiment. The immediate result was therefore a set of behaviors developed in a virtual room, rather than a machine already roaming his home.

He said coding and training the virtual duck took just a few hours, compared with the weeks he thought he might previously have needed to figure it out. That comparison was his own assessment of the work, not a timed comparison between two training methods.

The model watched the training, too

The setup combined GPT-6 Sol with Hermes, an agent-running framework from Nous Research, and an Nvidia DGX Spark computer. Nvidia’s Isaac Sim provided the simulation software. Knight described a workflow in which the language model participated in training, rather than stopping after it supplied a program for him to run.

Sol used Isaac Sim to feed instructions to another model that controlled the virtual robot. It then watched videos of the robot in action, assessed how well the behavior was working and suggested changes to the training. Those jobs connected instructions, observation and revision within the same experiment.

Knight credited that process with getting the duck to move around without toppling over. His account distinguishes the model overseeing training from the model controlling the robot: Sol supplied instructions and assessed the resulting behavior, while the control model carried out movements in the simulated environment.

A moonwalk became a backward jog

Not every requested movement emerged as intended. Knight initially asked Sol to teach the robot to moonwalk, giving it a GIF of Michael Jackson sliding across a stage as a reference. The model tried to copy the movement, but hours of effort produced what Knight described as a jaunty backward jog.

The vision system on my robot duck does not work well enough to be deployed in a real product.

Will Knight, writing in WIRED

The cat-butler job remains a plan

Knight’s intended household task goes further than walking toward a cat. He wants the robot to dispense treats automatically for his Bengal, Leona. One proposed design would put a 3D-printed storage chamber on the duck’s head, allowing it to release a treat by bowing. That mechanism was a possibility he described, not a completed part of the demonstration.

He expects the simulated behavior to transfer to the physical Microduck because he considers it relatively simple. That remains an expectation pending access to the hardware. The demonstrated work covers virtual movement and cat identification; the envisioned next step combines those behaviors with a physical treat dispenser for his pet.