6D Hybrid Pose Estimation in Cluttered Industrial Scenes for Robotic Grasping

Abstract

6D pose estimation is an important topic in robotic grasping and it has been widely studied. However, when pose estimation is applied in industry, problems such as occlusion, symmetries, and texture-less of objects made accurate pose estimation and robotic grasping challenging. In this paper, we present a hybrid pose estimation method that combines template-based estimation and convolutional neural network. Our method can handle industrial objects in clutter and quickly adapt to different objects on assembly line. Grasp angle can be specified manually by manufacturer with a single RGB image of the objects. Our proposed method outperforms template-based or end-to-end baselines in the experiments.