KEY POINTS
- OpenAI's latest models can build 3D simulations and guide robot training, according to WIRED and the researchers it interviewed.
- Researchers from Nvidia and UC Berkeley open sourced an agentic robot framework called Aspire this week.
- Reliable real-world robots remain hard to build, and the WIRED writer says his own robot's vision system is not product-ready.
AI models are making it much faster to train robots in simulation, according to a WIRED newsletter by Will Knight. He and several robotics researchers say OpenAI's newest models can handle work that used to demand weeks of effort and specialist skills.
Knight described training a virtual version of Microduck, a small robot from Pollen Robotics and Hugging Face, using OpenAI's GPT-6 Sol. In a few hours, he taught the simulated robot to walk around a room and spot bengal cats. He hopes the behavior will carry over to a physical unit, which he is still waiting for, and he plans to add a treat dispenser for his own cat.
For the project he also used Hermes, an agentic harness from Nous Research, and an Nvidia DGX Spark personal computer. According to WIRED, the model went beyond writing code. It used Nvidia's Isaac Sim software to pass instructions to a model controlling the robot, watched videos of the results to judge progress, and proposed changes to the training setup.
The process had limits. Knight said he asked the system to teach the robot to moonwalk from a GIF of Michael Jackson, and after hours of effort it produced only a jaunty backward jog.
Ken Goldberg, a professor at the University of California, Berkeley, said his lab tested OpenAI's Astra model after its release in September. A lab member gave it access to Nvidia's Newton tool and a video of a real robot, then asked for a simulation of the scene. Goldberg said that within an hour it had produced a full 3D simulation with the camera position, lighting, background and materials all set. He added that the model can also automate much of the tedious experiment-and-review work of simulated training.
Jim Fan, an Nvidia researcher who has worked with Goldberg, said the way AI now controls robots resembles how software agents operate. Language models can call on models trained to drive physical actions, known as visual language action models, much as they use programs on a computer. This week Fan, Goldberg and others released an agentic framework called Aspire as open source software. It lets a robot discover and use new skills, and in one example a GPT-6 model used two robot arms to solve a Rubik's cube.
Fan gave a keynote at the Conference on Robot Learning in Austin, Texas, this week. He showed a demo in which Astra controlled a robot arm and painted the Golden Gate Bridge with nearby paints and brushes.
WIRED notes that making robots work reliably is still very difficult and takes serious engineering skill, and Knight said his duck's vision system is not good enough for a real product. Even so, Goldberg said several industrial robotics and automation companies have shown interest, and Knight suggested the approach could affect manufacturing and e-commerce, and might bring a wave of hobbyist "vibe robotics".
