Efficiently Learning Manipulations by Selecting Structured Skill Representations

StructDiffusion: Language-Guided Creation of Physically-Valid Structures using Unseen Objects

“…Many of these do so explicitly. In particular, learning object skill preconditions is very useful for sequential manipulation, so some works look at predicting relationships in this context [29,27,30]. For example, SORNet [27] learns to predict relations between objects given a canonical image view of the objects; similarly a predictive model from image inputs is learned for capturing relationships [30].…”