Robots often learn skills in computer simulation before acting in the real world, since practising on hardware is slow and risky. But simulation cannot capture friction or how bouncily a ball rebounds, so skills learned inside often fail outside.
CUE is an agent, a program that acts on its own. It builds a view of the scene from camera images, guesses the missing physical properties, deliberately tries the shots its simulation predicts worst, and corrects the guesses from what actually happens.
On billiards, that cut error in predicting where balls end up by 72% against the best alternative, across 25 hidden physics settings, and it also ran on a real pool table.