Using Abstraction for Interpretable Robot Programs in Stochastic Domains