30 octobre 2026 à 14h
Abstract:
Model Predictive Control (MPC) computes control actions by solving an optimization problem online, which lets it handle constraints on states and inputs explicitly. Its guarantees, however, rely on an exact model. Tube MPC restores robustness for uncertain linear systems by predicting a sequence of sets, a tube, that contains every possible future trajectory. In this talk I will show that how this is done depends strongly on the geometry chosen for the tube and for the uncertainty: polytopes, ellipsoids, or a mix of the two lead to
different optimization problems, with different computational cost and different conservatism. I will focus on polytopic tubes with ellipsoidal uncertainty, which leads to a second-order cone program, and compare it with existing approaches. I will then discuss how conservatism can be reduced by learning the unknown parameters online with set-membership estimation, why ellipsoidal estimates raise issues of feasibility and excitation, and how a conditional update rule restores recursive feasibility and stability guarantees.
Francesco Licata, University of Lyon
Inria, plenary room
français
English