Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : May 10, 2017 - May 13, 2017
View-based approach is one of the methods to obtain environmental information by a camera. This method uses image data itself without object models or camera calibration, thus it can be applied to manipulation of various objects. In this paper, we studied a robot teaching method using view-based approach: view-based teaching/playback. We introduced deep learning with CNN (Convolutional Neural Networks) into view-based teaching/playback. In virtual environment, we succeeded in making a robot push an object to a goal position from teaching positions after multiple teaching demonstrations. We also succeeded in making the robot push it from some initial positions that were not identical to those in the demonstrations.
