One Policy to Dress Them All: Learning to Dress People with Diverse Poses and Garments

Textile Taxonomy and Classification Using Pulling and Twisting

“…Model-free learning is a class of methods that circumvents the challenges in state estimation and dynamics modeling. This technique has been successfully used to tackle several cloth manipulation tasks such as folding or flattening, by finding the best sequence of pick-and-place positions [10], [11], [12], [13], [14]. More recently, [15] proposed a model-free visual feedback policy to fold cloths in half, successfully adapting the manipulation trajectory to three real-world cloths.…”

“…Point clouds, on the other hand, demand less computational effort for perception tasks due to their unstructured nature and have shown success in assistive dressing tasks [13]. Nevertheless, in situations where clothes significantly self-occlude, point cloud representations become ambiguous as different layers of the cloth cannot be distinguished based solely on the observable set of points.…”

“…The reward function and the model architecture for this baseline correspond to the one chosen for our approach, where the critic architecture and hyperparameters match the ones of AdaFold. As the output of the critic has the same dimension as the number of input points, to extract the value from the output of the segmentation branch, we select the first point of the network output as similarly done in [13].…”