Physical edge detection in clothing items for robotic manipulation

Abstract

The physical edges of an object contain valuable information to recognize its shape and also to manipulate it. In the case of handling clothing items automatically, the physical edges give important clues to determine its type and shape as well as to find good grasping points for many manipulation tasks. In this paper, we propose a method to extract the garment's physical edges from the three-dimensional observation of the garment. First, we calculate depth edges as candidate pixels, and then use a Deep Convolutional Neural Network for pixel-wise classification of the edges. Experimental results obtained by using various towels of different sizes and softness show the robustness of this method. To demonstrate the usefulness of this detected physical edge information, we also conducted experiments of opening towels with a dual-arm robot.