To provide a feasible transitional solution from all-by-human driving style to fully autonomous driving style, this paper proposed concept and its control algorithm of a robust lane-keeping ‘co-pilot’ system. In this a semi-autonomous system, Learning based Model Predictive Control (LBMPC) theory is employed to improve system's performance in target state tracking accuracy and controller's robustness.
Firstly, an approximate LTI model which describes driver-vehicle-road closed-loop system is set up and real system's deviations from the LTI system resulted by uncertainties in the model are regarded as bounded disturbance. The LTI model and bounded disturbances make up a nominal model. Secondly, a time-varying model which is composed of LTI model and an ‘oracle’ component is designed to observe the possible disturbances numerically and it is online updated using Extended Kalman Filter (EKF). Thirdly, LBMPC method is used to decouple controller's optimal control performance and its robustness. Constraints are applied to states predicted by nominal system, while states predicted by oracle model are used for computation of cost function.
In the end, this paper presents simulation results obtained by Matlab/Simulink. Straight road and curved road are two typical driving conditions used for verification. The results show the conclusion that the EKF algorithm can estimate the un-modeled uncertain dynamics effectively and the proposed lane-keeping copilot based on the oracle model can successfully assist driver to stay in an expected driving zone.