A view on learning robust goal-conditioned value functions: Interplay between RL and MPC