Berkeley tests AI-written robot programs with success rates up to 99%

UC Berkeley researchers tested an approach that uses AI coding agents to write and refine robot programs, reaching success rates up to 99% in tasks where baseline imitation learning achieved 20%, IEEE Spectrum reported. Berkeley and Nvidia are announcing the approach at the Conference on Robot Learning. Agents divide tasks into parts, generate code using existing robotics tools and test it in simulation. They can also adjust simulations using video of real-world failures. The resulting code is readable by engineers. The results match human hand-engineered performance in the reported tests, but translating simulated behavior into reliable physical contact remains an unresolved challenge, according to Intrinsic chief science officer Torsten Kroeger.