Correcting Nominal Models with Learned Residuals for Aggressive Real-Time Control

Abstract

Car-like robots are mechanically simple and dynamically stable, making them well-suited for applications such as inspections, surveillance, and search-and-rescue operations. Given a model of the robot’s dynamics, model predictive control (MPC) plans optimal actions with specified constraints. The performance of MPC is limited by inaccuracies in the underlying system model - high fidelity models are complex and brittle, while simple models trade accuracy for computational tractability.
This work presents a hybrid control framework that corrects predictions from a low-dimensional kinematic model through a data-driven residual network, termed the Corrective Residual Network (CRN). CRN significantly reduces model errors compared to the nominal model. CRN is integrated with MPC in a proposed algorithm called Corrective Sequential Linear-Quadratic MPC (CSLQ-MPC). CSLQ-MPC demonstrates strong performance and robustness towards model mismatch, including in scenarios with incorrect model parameters given to the nominal model. CSLQ-MPC sees an improvement in cross-track error of 54.3% and an improvement in time taken of 14.3%.