How AI Is Teaching Robots to Learn New Skills on Their Own

Jim Fan, a researcher at Nvidia who has worked with Goldberg, compares the way AI now steers robots to how software agents operate. Large language models can call on models trained to govern physical movement—so-called vision-language-action (VLA) models—in much the same way they tap into different programs on a computer.

This week, Fan, Goldberg, and their colleagues open sourced an agentic framework for robots named Aspire, which lets a machine discover and adopt new skills on its own. In one demonstration, GPT-6 Astra worked out how to operate a pair of robotic arms to solve a Rubik’s cube.

“It picks it up, tries to twist it, but has the wrong angle,” Fan explains. “Then it reasons for a very, very long time, but it eventually solves it because it learns the skills over time.”

At the Conference on Robot Learning (CoRL) in Austin, Texas, this week, Fan delivered a keynote about a striking realization: AI models can automate much of the work required to get a robot to perform complex tasks. During his talk, he presented a demo in which Astra took control of a robotic arm and was instructed to paint the Golden Gate Bridge. Using the paint and brushes within reach, it completed the picture.

Building robots that operate reliably and efficiently remains a formidable challenge that demands deep engineering expertise. The vision system on my own robot duck, for instance, still isn’t refined enough to appear in a real product.

Even so, the speed and simplicity of training robots with AI could soon reshape manufacturing and e-commerce, and Goldberg notes that several industrial robotics and automation companies have expressed interest in the method.

Much as vibe coding did a few years back, we may soon witness breakthroughs in “vibe robotics” as hobbyists and engineers discover just how easily robots can be programmed and trained. And, if I can perfect my robotic helper, Leona may never have to go without treats again.

Source link