Bayesian Optimization for Sample-Efficient Policy Improvement in Robotic Manipulation