This paper presents a system which generates images of objects behind obstacles as if they are seen through the obstacles. For remote operation of robots such as disaster response robots, operational efficiency depends heavily on quality of visual information. Diminished reality, which has been proposed in the field of computer vision, is a technique which generates images in which as if users saw objects through obstacles. The proposed method uses a robot with an arm and two RGB-D sensors attached in front of and on the arm of the robot, acquires three-dimensional information of occluded areas and integrates the images to present see-through image, letting the sensor move along with the arm. In this method, past information of RGB-D images can easily be used to expand the field of view with low calculation cost, if the environment is static. The experiments with a manipulator robot show the ability to see through obstacles from arbitrary viewpoints in real time.