Safety-Guaranteed Learning-Predictive Control for Aggressive Autonomous Vehicle Maneuvers

Abstract

This paper seeks to securely maximize the maneuverability of autonomous vehicles using a nonlinear learning-predictive controller to enable safety-guaranteed aggressive vehicle maneuvers. In the design, a data-driven Gaussian process on polynomial basis (GPPB) method is used to improve vehicle dynamics model accuracy and sum-of-square (SOS) technique is employed to estimate the vehicle operation safety boundary. A nonlinear model predictive controller is designed to incorporate the GPPB-based model and the SOS-enhanced safety region estimation. Experimental and comparison results with other two benchmark control designs on a scaled vehicle testbed demonstrate the effectiveness of the proposed controller.