Affordance Transfer based on Self-Aligning Implicit Representations of Local Surfaces

Other conference contribution, 2022

Objects we interact with and manipulate often share similar parts, e.g. handles, that allow us to transfer our actions flexibly due to their shared functionality. This corresponds to affordances, i.e. set of action possibilities offered by the environment [1]. In this work, we propose to learn affordances associated with implicit models of local shapes shared across object categories. Our approach takes an expert grasp demon- stration on a given object, extracts the local geometry, and uses it as an anchor to align corresponding parts of objects from the same category. We show that the proposed implicit representation method can align objects within the same category under random pose perturbation. In addition, our general approach can align the local geometry to find grasp poses similar to the one demonstrated in the reference local shape. Finally, we show that we can identify the shared local geometry on novel objects from a different object category for affordance transfer.