Picking from clutter: An object segmentation method for robot grasping

Abstract

Picking in a unstructured environment is an important task for the further autonomy of the robot manipulation in real applications. A primary challenge for the task is to identify the object from the cluttered sensor readings. In this paper, a real time segmentation algorithm is proposed to partition the scene into objects using only depth and geometry information. We employ a graph to model the scene, in which the surfaces are regarded as nodes while the geometric relations between surfaces as edges. The relations are represented by the convexity and connectivity of the two neighbor surfaces. Upon the segmentation result, a measure was developed for robot grasping proposal suggestion. Our method has advantages over the RGB and learning based methods as it is robust against the illumination variation and does not require the collection of samples, thus achieving more convenient deployment. The method was evaluated on public datasets to validate its feasibility and effectiveness, demonstrating better performance compared to other depth information based image segmentation method. Besides, a real-world robot grasping experiment is conducted to investigate the possibility of on-site production. abstract environment.