There also might be a delta (variation of the activity) that optimizes the learning progress. My guess for that would be...change too little and you're inefficient, change too much and you are inefficient, too.
In the article they say that "Contextual variability can strengthen retention [15] and generalization of skills", but that they "chose to increase sensorimotor variability while maintaining constant the original learning context."
They also mention that "attributing errors to internal sources can strengthen learning [19, 20] and generalization [21] of motor behavior"
My guess for an autodidact would be how powerful internal attribution is compared to the change in context and exploration of the sensorimotor space. Of course, they're probably not discrete, and probably also impact one another.
Sounds like some really interesting possible implications, and possibilities for further research. Really makes me want to reread the article and go through some of the articles they cite
"Participants were unaware of this manipulation"
"We chose to manipulate sensorimotor variability so that participants were unaware of the fact that there was any change in the task"
[1] http://www.cell.com/cms/attachment/2045379445/2056784269/mmc...
So not only are the tasks similar, but they're so similar that participants weren't consciously aware that they were doing something different. Wacky
They also mention that "participants strengthened skill through the re-exploration of sensorimotor space."
I wonder how large an exploration of the sensorimotor space could be while generalizing/improving the original skill. (re batting and kendo). No idea, but interesting article either way