Learning-Based Model Predictive Control with Application in Robotic Trajectory Tracking

This paper studies the learning-based model predictive control problem for nonlinear systems with model uncertainties and control constraints. First, a prediction model is constructed offline. The prediction model is composed of a nominal model derived using the first principle with known parameters, and a learning model constructed via the LSTM network to account for model uncertainties and unknown disturbances. Then control input increments are optimized using an online model predictive controller with constraints. Simulation results for trajectory tracking with a robotic arm are presented to verify the robustness and feasibility of the proposed approach.