I Saw the Future of Robotics in a Virtual Duck That Chases My Cat
An OpenAI model helped me teach the Microduck to waddle and moonlight as a tiny cat butler.
OpenAI's GPT-6 Sol allowed me to train a virtual robot named Microduck to walk around a room and identify a Bengal cat. Although I'm eagerly awaiting a physical Microduck, the simulated behavior is relatively easy to replicate. I plan to train the robot to release a treat for my Bengal cat, Leona, using a 3D-printed storage chamber mounted on its head.
Coding and training the virtual robot took only a few hours, a remarkable feat compared to previous methods that could have taken weeks. This is because OpenAI's latest models excel at understanding and manipulating 3D representations of the physical world and translating them into useful code, actions, and robot training. Real roboticists, including University of California, Berkeley professor Ken Goldberg and his students, have also taken advantage of OpenAI's latest models to produce detailed and accurate simulations that help train different kinds of robots.
By leveraging Nvidia's Newton tool and an Nvidia DGX Spark personal computer, they were able to create a full 3D simulation in just an hour, complete with camera positioning, lighting, background, and material settings. Furthermore, OpenAI's newest model can automate much of the drudgery involved in training a robot in simulation by running experiments and examining the results.
This process allowed me to train the Microduck to walk without falling over. Initially, I attempted to teach the robot to moonwalk using a GIF of Michael Jackson, but it only resulted in a jaunty backward jog. Jim Fan, a researcher at Nvidia who has worked with Goldberg, explains that AI now controls robots in a similar manner to how software agents interact with a computer.
This week, Fan, Goldberg, and others open sourced an agentic framework for robots called Aspire, which enables a robot to discover and utilize new skills. For instance, GPT-6 Astra managed to teach a pair of robot arms to solve a Rubik's cube. At the Conference on Robot Learning (CoRL) in Austin, Texas, Fan discussed how AI models can automate much of the work necessary for robots to perform sophisticated tasks.
Despite the challenges in getting robots to work reliably and efficiently, the vision system on my robot duck is not yet advanced enough for real-world deployment. Nonetheless, the speed and ease of training robots with AI could significantly impact manufacturing and e-commerce, with industrial robotics and automation firms showing interest in this approach.
Written by urgent.news from Wired Business's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
