Imitation Learning of Nonlinear Model Predictive Control for Emergency Collision Avoidance
Seungtaek Kim, Kyoungseok Han, Seibum B. Choi
Abstract
This study proposes a control structure based on imitation learning (IL) of nonlinear model predictive control (NMPC) for vehicle collision avoidance systems. An NMPC was employed to achieve maximum collision avoidance ability by integrated steering and braking, then later imitated by a deep neural network (DNN) to satisfy real-time capability. Previous studies that imitate NMPC have proven its control performance and computation efficiency. However, there were limitations in applying to vehicle collision avoidance systems. Despite its dangerous situation, data set for imitation should be obtained by experiments using the controlled plant, and weaknesses in handling model parameters were shown. Therefore, this paper proposes a novel IL-based control structure suitable for collision avoidance systems that overcame such limitations by building a feedforward feedback structure so that the data set trained for imitation can be made offline and applying an input dimensionalization process to ensure robustness to parameter changes. CarSim-based human-vehicle interactive simulation experiments demonstrated that the proposed IL-based control structure had no issue applying the offline trained DNN in the simulation while showing robustness to parameter changes.