Learning-Based Safety Critical Model Predictive Control Using Stochastic Control Barrier Functions

摘要

This paper presents a learning-based safety-critical Model Predictive Control (MPC) design approach based on stochastic Control Barrier Functions (CBFs). To address the safety concerns and tackle model uncertainties in both the MPC and CBF, we first propose to use a parameterized stochastic CBF in the MPC scheme. We next devise a Reinforcement Learning (RL)-based algorithm based on the proposed stochastic CBF-MPC scheme to learn the approximate version of the proposed stochastic CBF for coping with an unknown CBF model, which cannot capture the correct structure of the CBF used in the real environment. To illustrate the performance of the proposed safety-critical control approach, we examine two test cases including trajectory tracking and path planning for a wheeled mobile robot.