Learning Self-Supervised Representations from Vision and Touch for Active Sliding Perception of Deformable Surfaces.