Improving safety in mixed traffic: A learning-based model predictive control for autonomous and human-driven vehicle platooning

Wang, J.; Jiang, Zhihao; Pant, Yash Vardhan · 2022 · OpenAlex-citations

DOI: 10.48550/arxiv.2211.04665

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the safety challenges inherent in mixed-traffic environments where autonomous vehicles (AVs) platoon alongside human-driven vehicles (HVs). The motivation stems from data indicating that 64.2% of accidents in mixed traffic involve HVs rear-ending AVs, a significant increase compared to conventional traffic. This risk arises because current control strategies often fail to adequately account for the unpredictable nature of human drivers. The authors aim to develop a control strategy that explicitly models HV uncertainty to enhance safety during longitudinal car-following scenarios, particularly in complex situations like emergency braking. To achieve this, the authors propose a novel hybrid modeling approach and a corresponding control strategy. First, they develop a human-driven vehicle model that combines a first-principles autoregressive with exogenous input (ARX) model, which captures deterministic human reaction delays, with a Gaussian process (GP) machine learning model. The GP component learns the discrepancies between the ARX predictions and actual driver behavior, thereby quantifying uncertainty. Data for training this model was collected from three drivers in a Unity simulator who were distracted by algebraic questions to simulate realistic variability. Second, they formulate a Gaussian Process-based Model Predictive Control (GP-MPC) strategy. Unlike baseline methods that use fixed safety margins, this controller utilizes the variance estimates from the GP model to define adaptive, probabilistic safety constraints (chance constraints) for the distance between the trailing HV and the leading AV. The results demonstrate significant improvements in both prediction accuracy and control performance. The hybrid ARX+GP model reduced the root mean square error (RMSE) of HV speed predictions by 35.64% compared to the ARX model alone, with average RMSE dropping from 1.88 to 1.21. In simulation studies, the GP-MPC strategy outperformed a baseline nominal MPC. The proposed method maintained larger minimum distances between vehicles while facilitating higher travel speeds, indicating superior safety and efficiency. The GP-MPC effectively leveraged the quantified uncertainty to adjust safety margins dynamically, proving more robust than methods relying on predefined bounds or deterministic models. The significance of this work lies in its integration of quantified human uncertainty directly into the control policy design. By using GP-derived variance as a constraint, the authors provide a method for AVs to adaptively manage safety risks posed by unpredictable human drivers. This approach bridges the gap between traditional parametric models and data-driven techniques, offering a more nuanced and safer framework for mixed-vehicle platooning. The study highlights the potential of learning-based MPC to handle stochastic human behaviors, suggesting a pathway for safer integration of AVs into existing traffic systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness