Using predicted future visual frames to plan robot actions without explicit physics simulation or state representations.