VIBR: Learning View-Invariant Value Functions for Robust Visual Control