Developer trains virtual robot duck to chase cat with AI

Virtual robot tracks feline target

The creator utilized OpenAI's GPT-6 Sol to instruct a digital version of the Microduck from Pollen Robotics. Within hours, this virtual agent mastered walking through a room and spotting any Bengal cats in the vicinity. The system currently simulates behavior that will eventually transfer to a physical unit once it ships. A video clip shows the robot waddling toward the target feline while displaying its eye view. The developer plans for the real robot to dispense treats by bowing its head to release them.

AI tools automate simulation setup

Training required an agentic harness from Nous Research called Hermes alongside a Nvidia DGX Spark computer. OpenAI released Astra in September, allowing University of California Berkeley professor Ken Goldberg to test 3D rendering capabilities. His lab member gave Astra access to an Nvidia tool called Newton to create detailed simulations quickly. Within one hour, the system produced a full 3D simulation with camera positions and lighting parameters set. These tools automate drudgery by running experiments and examining results to suggest tweaks to training regimes.

Initial moonwalk attempt failed

The developer initially tried getting GPT-6 Sol to teach the robot a moonwalk using a Michael Jackson GIF. After hours of effort, the model only resulted in a jaunty backward jog instead of sliding across the stage. Jim Fan from Nvidia notes that AI control mirrors how software agents function in current systems. Large language models can now use trained models to manage complex robotic behaviors effectively. Real roboticists are using these latest models to achieve amazing results in their own projects.