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

The writer trained a virtual Microduck robot to walk around a room and recognise Bengal cats, using OpenAI’s GPT-6 Sol. The experiment suggests that newer AI models can do more than write robot code: they can help set up simulations, assess results and adjust training, potentially making robotics work faster and more accessible.

The writer says the virtual robot was trained in a few hours with Nvidia Isaac Sim, a Nous Research agentic harness and a DGX Spark computer, and learned to walk without falling over. A separate Berkeley lab demonstration reportedly produced a detailed robot simulation in about an hour. Researchers have also open-sourced Aspire, a framework for robots to discover and use skills; one example involved Astra learning to solve a Rubik’s cube with robot arms. The article notes that the Microduck is still virtual, so the writer’s plan to have it dispense treats to a pet cat remains prospective.