Long-Horizon Vehicle Motion Planning and Control Through Serially Cascaded Model Complexity
Vincent A. Laurense, J. Christian Gerdes
Abstract
The computational burden of nonlinear model predictive control (NMPC) often limits its use to short planning horizons, simple systems with slow dynamics, offline applications, or approximations of the optimal control problem. This article introduces a novel concept for NMPC, to help enable real-time integrated motion planning and control with a long planning horizon for automated vehicles. The proposed framework cascades plant models of different levels of complexity within a single planning horizon, in a single optimization problem. Leveraging the receding nature of MPC, a high-fidelity plant model in the first part of the planning horizon continuously provides a high quality of control, while the planning horizon is extended significantly at low computational cost with a lower fidelity model. Cascading the model complexity serially in a single planning horizon, rather than in different control loops, avoids infeasible reference trajectories between control loops. The concept is successfully validated with real-time motion planning and control of a full-scale automated race car, featuring combined lateral and longitudinal control and operating the vehicle near the limits of tire-road friction. The framework is deployed with open-source numerical optimization tools. In the real-world experiment, the proposed design both better approaches the optimal minimum-time solution and has a lower median solve time compared to a benchmark architecture with a single-vehicle model and a necessarily shorter planning horizon.