RGB-D-Based Robotic Grasping in Fusion Application Environments

摘要

Although deep neural network (DNN)-based robot grasping has come a long way, the uncertainty of predicted results has prevented DNN-based approaches from meeting the stringent requirements of some industrial scenarios. To prevent these uncertainties from affecting the behavior of the robot, we break down the whole process into instance segmentation, clustering and planar extraction, which means we add some traditional approaches between the output of the instance segmentation network and the final control decision. We have experimented with challenging environments, and the results show that our approach can cope well with the challenging environment and achieve more stable and superior results than end-to-end grasping networks.