Training with one type of a visual stimulus usually improves performance. When observers train with two or more stimulus types presented in random order (so-called roving), performance improves for certain stimulus types but not for others. To understand when roving hinders perceptual learning, we conducted four experiments, using four different pairs of stimulus types. Performance improved when a bisection and a Vernier stimulus were randomly interleaved. However, no learning was found when we randomly interleaved this bisection stimulus with another one being twice as long. Our results suggest that roving impedes perceptual learning when the stimulus types are perceptually clearly distinct but, still, excite overlapping neural populations