Provably Safe Deep Learning-Based Offset-Free Model Predictive Control and its Application to Safety-Critical Autonomous Driving

Offset-free model predictive control (MPC) provides a useful means for controlling systems with uncertainties and constraints, but suffers from the heavy computational burden of repeatedly solving an optimization problem in real time. Such computational issue precludes the possibility of its application in systems requiring high realtime requirements, such as autonomous driving system. To address this problem, we develop a provably safe deep learning-based offset-free MPC framework. Based on the nominal offset-free MPC, the proposed MPC not only reserves the ability of disturbance rejection, but also leverages deep neural networks for approximating the explicit MPC solution to greatly reduce online computational time. Furthermore, a gauge map is used to guarantee the satisfaction of safe constraints. The proposed MPC is used in trajectory tracking control for smart autonomous driving. The simulation results show that the proposed MPC is an order of magnitude faster than the nominal offset-free MPC in safety-critical systems.