Towards Provably Safe And Robust Learning-Enabled Systems