Data-Driven Incremental Model Predictive Control for Robot Manipulators

Abstract

Model predictive control (MPC) is one of the few control frameworks allowing to systematically integrate input and/or state constraints to realize safe control. Nonetheless, traditional MPC requires a full model of the system dynamics and model mismatch degrades performance. In this article, a data-driven MPC scheme is developed for robot manipulators to reduce dependencies on system models and also design parameters. First, an incremental MPC is designed using an approximated model from time-delay estimation (TDE) that allows for prediction based on recent information from sensors instead of a full model. Then, a data-driven parameter estimation method updates the TDE parameter online to reduce dependence on parameters and improve accuracy of this approximated model. A recursive least squares algorithm is used and equipped with a novel strategy to adapt the forgetting factor based on the variation of the identified parameters. The resulting data-driven MPC allows for efficient implementation and we demonstrate its superior tracking performance in experiments with a 3-DoF robot manipulator.