Robots are notoriously bad at the things humans find easy, like picking up an oddly shaped object without dropping it. New research suggests a way around that problem: let the robots improve themselves with help from AI coding agents.
According to The Decoder, researchers from Nvidia, Carnegie Mellon University, and UC Berkeley used AI coding agents to teach robots dexterous grasping in the real world. Rather than relying solely on humans to hand-tune every behavior, the approach puts AI coding agents in the loop to help the machines train themselves.
The results are striking. The Decoder reports that a fleet of eight robots reached up to 99 percent success on tricky tasks. That kind of reliability matters because grasping is one of the long-standing stumbling blocks in robotics, where even small failure rates can make a system useless for real work.
The collaboration is also notable for who is behind it. Nvidia, best known for the chips powering the AI boom, is teaming with two of the top US research universities, signaling continued investment in robotics as a frontier for artificial intelligence.
The details in the available reporting are limited, and the work comes from a research setting rather than a shipping product. Still, the headline claim — robots that train themselves through AI coding agents — points to a shift in how machines might be taught.
Why it matters: if AI agents can help robots learn delicate physical skills on their own, the slow, hands-on work of programming machines for the real world could get dramatically faster.