They're both a superset of the problem space: safely navigate from A to B.

And broadly speaking, with more varied training data you can achieve better out of sample results.