Quality Assessment of Robotic Grasping Using Machine Regularized Leaning Algorithms

Abstract

Scientists would be delighted in providing robots with complicated skills that humans have. Grasping is considered one of the skills that the robots need to accomplish. Currently, robots with the use of deep learning are doing well in moving items through grasping. However, robots would never be able to improve their grasping without a measurement that quantifies how well the robot is grasping an object. The main goal of this letter is to evaluate the grasp robustness of a three-fingered manipulator based on a set of features (position, velocity, and effort) measured at each joint of the robotic hand. Results yield 93.4% accuracy of predicting whether a grasp is stable or unstable without the use of tactile sensors, without any prior knowledge of the gripped object, and without defining the end-effector geometrical properties. The ultimate objective behind this letter is to enhance the performance of inexpensive tactile sensors, yet deliver better grasping quality. This work can be ultimately used in robotics to save time from regrasping, to get a safer workplace, and to achieve a better manufacturing process.