My House, My Rules: Learning Tidying Preferences with Graph Neural Networks

Robots that arrange household objects should do so according to the user's\npreferences, which are inherently subjective and difficult to model. We present\nNeatNet: a novel Variational Autoencoder architecture using Graph Neural\nNetwork layers, which can extract a low-dimensional latent preference vector\nfrom a user by observing how they arrange scenes. Given any set of objects,\nthis vector can then be used to generate an arrangement which is tailored to\nthat user's spatial preferences, with word embeddings used for generalisation\nto new objects. We develop a tidying simulator to gather rearrangement examples\nfrom 75 users, and demonstrate empirically that our method consistently\nproduces neat and personalised arrangements across a variety of rearrangement\nscenarios.\n