Two-Layer MPC Architecture for Efficient Mixed-Integer-Informed Obstacle Avoidance in Real-Time

Abstract

Safe and efficient obstacle avoidance in complex traffic situations is a major challenge for real-time motion control of connected and automated vehicles (CAVs). Limited processing power leads to a trade-off between real-time capability and maneuver efficiency, especially for trajectory planning in highly dynamic traffic environments like urban intersections. Addressing this problem, we propose a novel two-layer model predictive control (MPC) architecture utilizing a differentially flat representation of the kinematic single-track vehicle model for optimal control. While a real-time capable quadratic programming-based MPC ensures local obstacle avoidance at every time step, its problem formulation is asynchronously updated by the globally optimal solution of a computationally more expensive mixed-integer MPC formulation. Both optimization problems are computed in parallel and incorporate position predictions of surrounding traffic participants available via vehicle-to-everything (V2X) communication. Collision-free and efficient obstacle avoidance in real time under realistic model errors is validated via high-fidelity co-simulations of typical urban intersection and highway scenarios with the traffic simulator CARLA.