Enhancing Safety in Mixed Traffic: Learning-Based Modeling and Efficient Control of Autonomous and Human-Driven Vehicles

In light of the rising prevalence of autonomous vehicles (AVs) on public roads, the development of effective and efficient control strategies for AVs is paramount, especially in the face of the uncertain behaviors of human-driven vehicles (HVs). In this paper, we present an advanced learning-based method for HV modeling. Our approach combines a firstprinciples model with a Gaussian process (GP) learning component, resulting in improved velocity prediction accuracy and a quantifiable measure of uncertainty. The novel HV model was estimated and evaluated using real-world data from field experiments. The derived HV model was subsequently employed to devise a GP-based modelpredictive control (GP-MPC) strategy with the aim of boosting safety in mixed vehicle platoons by incorporating uncertainty evaluation into the distance constraints. Simulated studies were conducted to benchmark our GP-MPC strategy against a standard model predictive control (MPC) that solely depends on the first-principles model. Our analysis revealed that the proposed GP-MPC methodology provides superior safe distancing consistency and promotes more efficient vehicular behaviors (notably higher transit speeds) within the mixed platoon. Through integrating a sparse GP method in HV modeling and implementing a dynamic GP predictive scheme within the MPC framework, we achieved a considerable reduction in the average computation time for the GP-MPC to a mere 5% increase compared to the standard MPC. This marks a significant advancement from our previous work, being approximately 100 times faster than the model that did not use these approximations.