LSTM Neural Network-based Predictive Control for a Robotic Manipulator

In this paper, a Predictive Control based on LSTM Neural Network and Differential Evolution optimization performs as a high-accuracy control of a Robotic Manipulator. Such system dynamics are known as highly non-linear systems, where multiple input and outputs are involved. Therefore, Model Predictive Control was selected as a regulator system to follow a complex trajectory given by a simulated problem. Based on simulated data, we trained a Neural predictor as an approximation of each robot-joint dynamics, where the controller computes an optimal signal for a reference-tracker problem. We validate our claim with a numeric simulation where a mechanical model is employed. Our results show an increase in precision and vibration reduction while demonstrating the feasibility of a Predictive control law with Differential Evolution optimization in this scenario.