Exploiting Model Uncertainty Estimates for Safe Dynamic Control Learning