Towards clothes hanging via cloth simulation and deep convolutional networks

Abstract

People spend several hours a week doing laundry, with hanging clothes being one of the laundry tasks to be performed. Nevertheless, deformable object manipulation still proves to be a challenge for most robotic systems, due to the extremely large number of internal degrees of freedom of a piece of clothing and its chaotic nature. This work presents a step towards automated robot clothes hanging by modeling the dynamics of the hanging task via deep convolutional models. Two models are developed to address two different problems: determining if the garment will hang or not (classification), and estimating the future garment location in space (regression). Both models have been trained with a synthetic dataset formed by 15k examples generated though a dynamic simulation of a deformable object. Experiments show that the deep convolutional models presented perform better than a human expert, and that future predictions are largely influenced by time, with uncertainty influencing directly the accuracy of the predictions.