You have to figure out messy real life tasks anyway (ambiguous reward signal, inconsistencies, variables outside the observation).

Having actual data just makes it more feasible.

very very easily