I tried this once in the past and failed, but I was also doing it wrong, methinks.
My suspicion is that if you're doing it yourself, you'll need to do quite a bit of experimenting.
One thing for sure is that the more weight of the robot is on the ball, the hard it is to spin that ball, so weight probably needs to be supported by other parts, for example 3 or 4 rotating casters positioned around the central dyson ball.
Ultimately, I'm under the impression that it's a very inefficient method of locomotion.
In CMU's case, the small footprint of the robot is probably the main benefit. The thing that helps them accomplish it, too, is probably high tolerances, and strong motors.
Hope that helps and wasn't too incoherent