Current students


Section: Systems and Control

Major Research topic:
Continual Learning Model Predictive Control: theory and applications

Iterative Learning Control (ILC) employs the information about the error made in past trials of repetitive tasks in order to iteratively adjust a parametric control trajectory. In recent years, thanks to the availability of computational power, the ILC paradigm has been applied also to model predictive controllers. The resulting Iterative Learning MPC formulation shows a number of advantages: 1) it can easily deal with both linear and nonlinear systems 2) several control problem formulations are allowed (e.g., Economic MPC) 3) explicit constraints can be embedded.
In the past 10 years, different formulations of Iterative Learning MPC have been proposed. The Learning MPC proposed by F. Borrelli and coauthors is a one-shot approach to iteratively find the optimal trajectory and follow it. The Reinforcement Learning Model Predictive Control by S. Gros and coauthors iteratively adjust the MPC parameters in order to ensure stability and tracking performance even under the assumptions of relevant modeling errors and stochastic environment.

This work aims to combine the results obtained in existing literature into a unified framework to deal with both the problems of generating an optimal reference trajectory and tuning the MPC parameters to account for model uncertainty and disturbances.
This unified framework is further extended by expanding current Iterative Learning MPC formulations into a new framework for model-based learning control called Continual Learning Model Predictive Control. In this new approach, the MPC parameters are optimized sequentially, but the tasks may be potentially different along the iterations, within a transfer learning rationale mutuated from the machine learning literature. It follows that the resulting predictive controller is aimed both to deal with new tasks (forward transfer) and improve on old tasks (backward transfer).
The effectiveness of the proposed method is assessed on different applications, with a particular focus on both collaborative and mobile robotics.