Chaochen Gu, Qi Feng, Changsheng Lu, Changjian Gu, Xunjin Wu, and Kaijie Wu
Active vision, purposive perception, viewpoint estimation, pose correction, automated robotic manipulation
Robot vision has been widely applied in industrial robot manipulation tasks [1]. There is a tendency to equip robots with the capability of autonomous task planning using vision techniques in unstructured environments. This paper presents an active vision framework that is feasible to manage workpiece pose estimation in industrial application. The framework mainly includes two stages: 1) active viewpoint transfer and 2) accurate pose correction. In order to make the end-effector of an eye-in-hand robot automatically transfer from initial observed viewpoint, which exhibits occlusions or lacks of feature, to a largely different purposive viewpoint that is better for perception, an effective strategy of fast active viewpoint transfer is proposed. However, as the obtained viewpoints is still coarse and inadequate for precise robot manipulation, we further develop a pose correction method based on semantic features, which could address the challenges of feature extraction and registration on texture-less objects. With the proposed framework, extensive experiments were performed and the corresponding results reveals its potential in active vision.
Important Links:
Go Back
