Robotized Grasp: Grasp manipulation using Evolutionary Computing

Abstract

Grasping is a complex, multifaceted problem which requires substantial learning by humans. Executing the same by manipulating robot poses several challenges and so far, mostly those challenges are solved using planning based approaches. Recently, learning based strategies are getting momentum due to its inherent strength in handling dynamic changes in the manipulating environment. In this work, we present learning based technique for placing the robot hand correctly on the object to be grasped. The mapping of 2-D image space to the 3-D manipulators tip pose estimation has been learned using three machine learning approaches namely, Linear Regression (LR), Pseudo Inverse (PI) and Genetic Algorithm (GA). For depth calculation of object before grasping, we have used Infrared Range Sensor (Analog IO) of Anukul (Baxter) robot to find the distance between graspable object and left hand end-effector (EE), since out of two hands of Baxter for experimentation we have used left hand only, although the other hand can also be used without loss of generality. We performed rigorous experiments using state of the art manipulating robot Anukul available in the robotics and artificial intelligence laboratory of IIIT-Allahabad. The results show using evolutionary computing manipulator position for the graspable objects can be learned very accurately compared to other methods, provided sufficient training data are made available.