Action planning for retrieving one item from a cluster of connected items using affordance-based prediction

This paper describes a method for retrieving a target object to which direct access is blocked by other objects. We specifically consider the case where these objects are physically connected via deformable stems and branches. We highlight the importance of affordance-based prediction for successful object retrieval in such situations. We propose a framework for generating action sequences using a graph neural network. Through simulation-based experiments, we confirm the functionality of predictions and action sequences generated by the proposed method, and report simplified fruit harvesting experiments using a physical robot.