Tactile sensors can provide detailed contact in-formation that can facilitate robots to perform dexterous, in-hand manipulation tasks. One of the primitive but important tasks is surface following that is a key feature for robots while exploring unknown environments or workspace of inaccurate modeling. In this paper, we propose a novel end-to-end learning strategy, by directly mapping the raw tactile data acquired from a GelSight tactile sensor to the motion of the robot end-effector.Experiments on a KUKA youBot platform equipped with theGelSight sensor show that 80% of the actions generated by a fully trained SFDQN model are proper surface following actions; the autonomous surface following test also indicates that the proposed solution works well on a test surface.