This paper explores how representations can emerge from networks of local anticipations acquired through interaction with the environment. The nature and relationships between such representations are studied in the context of the Tetris game, where the various elements of the board are differentiated from the sensorimotor flow. Spatiotemporal regularities are learned by predicting the changes in the visual flow resulting from gaze shifts and keyboard actions. While general predictions about saccades across non-deformable objects most often apply and structure space, specific predictions about rotations of larger block sets help differentiating the various tetrominoes (i.e. game pieces).