Improving and evaluating robotic garment unfolding: A garment-agnostic approach

Current approaches for robotic garment folding require a full view of an extended garment, in order to successfully apply a model-based folding sequence. In this paper, we present a garment-agnostic algorithm that requires no model to unfold clothes and works using only depth data. Once the garment is unfolded, state of the art approaches for folding may be applied. The algorithm presented is divided into 3 main stages. First, a Segmentation stage extracts the garment data from the background, and approximates its contour into a polygon. Then, a Clustering stage groups regions of similar height within the garment, corresponding to different overlapped regions. Finally, a Pick and Place Points stage finds the most suitable points for grasping and releasing the garment for the unfolding process, based on a bumpiness value defined as the accumulated difference in height along selected candidate paths. Experiments for evaluation of the vision algorithm have been performed over a dataset of 30 samples from a total of 6 different garment categories with one and two folds. The whole unfolding algorithm has also been validated through experiments with an industrial robot platform over a subset of the dataset garments.