Approximate Robust Tube Nonlinear Model Predictive Control for Vehicle Collision Avoidance

The key to vehicle collision avoidance is achieving optimal avoidance performance with a reasonable computational load for real-time applications. To address these requirements, this study applies a novel approach by designing a robust tube nonlinear model predictive controller (RTNMPC) and approximating it to a neural network, thereby ensuring both optimal collision avoidance performance and realtime capability. The RTNMPC optimally controls the vehicle's steering and differential braking forces to guide it to a safe lane, minimizing the avoidance trajectory area. Tightened tire grip constraints were applied to robustly maintain vehicle maneuverability under system uncertainties and approximation errors in the neural network controller. Grip constraints were further relaxed by introducing a practical constraint tightening approach with an input saturation process based on tire grip usage. Consequently, the proposed collision avoidance system achieved both greater collision avoidance results with the lowest computational load compared to the baselines in CarSim simulations.