Robot Control by Least-Squares Policy Iteration with Geodesic Gaussian Kernels