An efficient robust model predictive control (MPC) strategy using augmented ellipsoidal invariant sets is enhanced by introducing more general dynamics in the future control moves (referred to as prediction dynamics) than simple time recession. This allows significant enlargements of the region of attraction for a give control horizon. As control horizons down to a single time step can be used, this implies online efficiency gains without decreasing performance. The procedure renders the offline problem a bilinear matrix inequality (BMI), thus algorithms for solving the non-convex offline problem is discussed.