Inverse kinematic solver based on machine learning sequential procedure for robotic applications

Abstract

This paper presents an inverse kinematic solver for a robotic arm based on an artificial neural network, ANN, architecture. The motion of the robotic arm is controlled by the kinematics of the ANN. The novelty of the proposed method is the validation using a proprietary robot of a novel procedure that applies three networks in a sequential mode to predict one joint value at a time. The inclusion of the genetic algorithm in the ANN definition and the adoption of sequential technique significantly reduced the manual settings and increased the obtained accuracy with respect to the traditional network deployment. The simulated outcomes proved the efficacy of the proposed approach in robotic motion control. The final architecture has three hidden layers: {40 (tansig), 35 (tansig), 30 (tansig)}. The resultant MSE in joint space is close to 3.235*10-4 rad2 and 0.1318mm2 in Cartesian space for the testing dataset. The maximum trajectory error for the validation curves, a planar circle and a spatial spring, is lower than 0.27mm for each axis.