Uncertainty-Aware Modeling of Learned Human Driver Steering Behaviors on High-Difficulty Maneuvers: Comparing BNNs and GPs

Abstract

We present a comparative analysis of two Bayesian approaches to representing internal models in human driver steering control: Gaussian Processes (GPs) and Bayesian Neural Networks (BNNs). We apply these methods within a recently proposed driver model that combines an uncertainty-aware learned internal model of the vehicle dynamics with a Model Predictive Control (MPC) framework, enabling representation of a range of driver behaviors. The GP and BNN approaches are evaluated with a focus on their ability to represent observed human behaviors, along with computational efficiency, prediction accuracy, and robustness. The GP model is shown to be particularly effective and robust in low-data regimes, providing interpretable uncertainty estimates leading to control that aligns well with human behavior observed during repeated driving maneuvers. BNNs, while computationally intensive to train, offer superior scalability and flexibility when dealing with high-dimensional systems or more complex and nonlinear dynamics, though predictions and learning are less interpretable.