Dual MPC for active learning of nonparametric uncertainties

This paper presents a dual model predictive control (MPC) framework for nonlinear systems with nonparametric uncertainties. We propose a dual MPC formulation using a Gaussian process model of the residual dynamics, which leverages the posterior state covariance to explicitly capture the dual effect and balance identification and control. By incorporating the GP covariance into the MPC objective, the resulting controller enables active learning, steering the system toward informative regions while balancing regulation to the desired reference. Robust constraint satisfaction of the novel dual MPC is established through a contingency plan. We demonstrate the proposed dual MPC on a nonlinear mechanical system and a robot navigation problem.