Robots are expected to perform in real life environment. We are focusing on wiping daily supplies. However, making computational model for every object has limitation. We propose a deep learning model for a robot that can wipe objects according to the shape by itself. First, let the robot experience wiping motions by direct teaching. Then, the force, tactile sense, image information and joint information are collected and input to neural network for training. Finally, we check the robot's ability to wipe other untrained objects which have more complicated shapes.