Automatic scene parsing for generic object descriptions using shape primitives

Abstract

Autonomous robots need to generate complete 3D models from a limited range of view when trying to manipulate objects for which no model is known a priori. This can be achieved by detecting symmetrical parts of an object, thus, creating an estimate of the invisible back sides. These symmetrical parts are typically modeled as primitive shapes (cylinders, spheres, cones, etc.), and fitted to noisy sensor data using sample consensus methods. This has the advantage that feasible grasps can be chosen from a precomputed set based on the estimated model, instead of a time-consuming random sampling approach. This article will look at fitting such analytic models to noisy 3D data in the context of robotic manipulation. State of the art methods from the Point Cloud Library (PCL) were extended to include additional relevant shapes (e.g. boxes), constraints (e.g. on size and orientation), and to consider additional information like knowledge about free space or proprioceptive information. A core part of the approach is the development of a scene parsing language, that allows for an easy-to-use pipeline specification during autonomous operation as well as shared-autonomy scenarios. Experiments will be presented based on scenes captured using an Xtion sensor.