Robotic Grasping Using Semantic Segmentation and Primitive Geometric Model Based 3D Pose Estimation

Due to the rapid development of hardware and software technologies, the machine automation has greatly improved in the past few decades. Robotic manipulators have been used in many factories for fully automated production. Compared with traditional sensor-based robot control, visual servo control has several advantages: high flexibility and precision, and robustness to calibration errors. In this paper, we present an eye-in-hand robotic arm for object recognition and grasping applications. The 2D images and depth information are acquired with an RGB-D camera for instance segmentation using convolutional neural network. The object pose estimation is achieved by ICP using primitive geometric models. Experiments carried out with different categories of objects in the real scene environment have demonstrated the feasibility of the proposed technique.