Scanning and Affordance Segmentation of Glass and Plastic Bottles

Abstract

Glass objects and objects made from translucent plastics are difficult to scan using ordinary 3D scanning techniques as they violate the Lambertian assumption that most 3D scanning technology relies on. We first present a method using Neural Radiance Fields to create meshes of transparent and translucent objects. After scanning the object it is useful to identify parts of the objects that can be manipulated by a robot. For example, to grasp a bottle we should identify a grasping point on the main body of the bottle, but to open the bottle we need to identify a grasping point on the cap of the bottle. In this paper we discuss scanning results of translucent bottles using Neural Scanning and we discuss an algorithm for segmenting the bottles into task specific parts.