Ensemble Model Predictive Control: Learning and Efficient Robust Control of Uncertain Dynamical Systems

Abstract

This paper presents a new robust Model Predictive Control (MPC) formulation using Ensemble Kalman Sampler to learn the parametric uncertainty of the dynamical model used for control design. It derives a polytopic model of uncertainty from data, and then uses the model to compute robust optimal trajectories while respecting input bounds and state constraints. Using linear dynamics the resulting controller can be written as a quadratic program, and under some assumptions we guarantee the constraint set forward invariant using the uncertainty model derived from data. We then describe extensions of the technique to non-linear autonomous and control-affine dynamics using Koopman spectral methods. Simulation studies of fast multirotor vertical landing illustrate the method.