Each step, Adam nudges every weight by about the learning rate. If the nudge doesn't get a weight halfway to the next grid value, it rounds back.