To evaluate the generalization of referring image segmentation (RIS) in the context of human-robot interaction, we generate referring expressions for a subset of images from GraspNet using Shikra.