Instance segmentation has developed for a long time, and many methods can produce accurate and sharp result in cluttered scene and deal with novel objects without pre-knowledge. We think the study of segmentation should consider its application. In this work, we propose policy-oriented Instance Segmentation for Ambidextrous Robot Picking. The robot has a Parallel- jaw gripper and a suction cup, and the two grippers complement each other.

We provide a dataset for ambidextrous robot picking that contains 6k synthetic scenes and 100 real scenes (dataset format is described in the paper).

You can download our dataset here.

The choosed ground truth lables for training our POIS network are available here.

E-mail: pois.casia@gmail.com

Authors: Guangyun Xu*, Yi Tao*, Bowen Jiang*, Peng Wang, Jun Zhong