An adaptive robotic grasping with a 2-finger gripper based on deep learning network

In this paper, an adaptive and versatile robotic grasping system is presented that is able to manipulate manufactured objects in production factories with a 2-finger gripper. A pick and place scenario based on deep learning framework is implemented and is achieved based on the following main steps: detection of the manufactured objects in the global scene observed by a first RGB-D camera using a first deep learning network, estimation of the object pose using 2D bounding box coordinates and depth information, motion of the arm above the object in an approach pose using Kinematics and Dynamics Library (KDL), recognition of the object’s face using a second deep learning network and information coming from a second RGB-D camera setup on the arm wrist, decision on the optimal grasping mode (opening or closing the fingers), execution of the grasping action. The developed system is validated practically by experiments in real world settings using a mobile manipulator platform consisting of 6 DoF robot arm with a 2-finger gripper setup on a mobile robot equipped by two RGB-D cameras.