Is it Necessary to Solve the Redundancy Problem when Learning the Inverse Kinematics of a Robotic Arm

Abstract

The aim of the this study was to investigate if it is possible to learn the inverse kinematics of a redundant arm without explicitly solving the redundancy problem. A 2D twojoint arm was simulated and a feed forward artificial neural network was trained to associate a desired endpoint and a current joint angle configuration with a correct joint angle displacement. Training examples were collected by allowing the arm to move around randomly. The model performed well, correctly reaching the desired endpoint 96% of the time. This shows that a system capabel of learning inverse kinematics do not need to explicitly solve the redundacy problem given the right kind of training examples.