Bayesian Optimization Meets Riemannian Manifolds in Robot Learning