An AI agent reportedly builds a robot-task simulation from phone video in under an hour

A post describes RoboSimStudio as an open toolkit built on Newton by researchers from UC Berkeley and NVIDIA Robotics. It says Claude and GPT-6 Astra each built a simulation of a two-arm wiping task from smartphone video, without CAD or calibrated cameras; one finished in 56 minutes. When the real robot’s bar support shifted instead of staying fixed, an agent changed the model to let it move. The revised simulation reproduced the tipping.