Simulation-Aided Policy Tuning for Black-Box Robot Learning