Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 150307

150307 - Human-Automated Vehicle Interactions: Voluntary Driver Intervention 

The current, and predictably the future, traffic environment is saturated with vehicles equipped with various automation features (SAE levels 2--4). Most notably, Adaptive Cruise Control (ACC) has enjoyed large market deployment. Such systems have been extensively studied in the literature, noting their benefits for safety, implications for traffic operations, etc.. A critical, yet often overlooked, implication of these technologies is the nuances of human-machine interactions and their impacts on the observed performance and benefits of such technologies. One of those critical interactions is the human control takeovers (often referred to as control disengagement). Yet such disengagement from automation and control takeover by human drivers is ubiquitous. We distinguish here between two kinds of control takeovers: (1) requested by the automation system, and (2) voluntary takeover by humans even when not critical. In the first, the control takeover is often initiated and requested by the automation system under various scenarios, including safety-critical ones. For such scenarios, a wealth of human factor studies have investigated how efficiently and safely human drivers can assume control of the vehicle. However, case (2) is less studied. Recent empirical evidence from naturalistic driving data found that voluntary takeover by human drivers (when not prompted by the system) is common.  It is estimated that a voluntary takeover occurs on average once every four miles. Voluntary takeover is attributed to driver trust in automation, or lack thereof. Particularly, studies suggest that driver's trust in automation erodes when the automated driving style is dissimilar to the human driving style. 

Regardless of who initiates takeover, control transition could spell trouble for traffic flow stability. Intentional control takeover stems from the necessity or desire to substantially change the course of driving. Thus it can involve a sudden change in driving behavior that could propagate through the traffic stream and cause major traffic disturbances. Notably, it is well-documented that even a subtle deceleration-acceleration movement by a vehicle can eventually develop into a full-blown stop-and-go disturbance in congested, human-driven vehicular traffic. Traffic disturbances instigated during control transition could be more severe and thus lead to greater traffic instability. 

Clearly, safety comes first, and efficient control transition is sometimes necessary. However, as studies have shown, traffic instability can also have negative safety consequences (e.g., rear-end collisions). Thus, it is desirable to reduce unnecessary, voluntary driver intervention. A solution to this would be to align the automated driving style with the human driver’s liking. The current ACC systems allow for some customization by enabling the human driver to adjust specific parameters, such as time headway or speed, to their driving preference. However, the alignment between human preference and automated driving style remains largely an afterthought, and takeover implications are yet to be fully studied and designed for.

This paper is then concerned with how voluntary driver intervention affects traffic and how to mitigate it. Towards this end, the objectives of this study are two-fold: (1) characterize voluntary interventions, vehicle kinematics during interventions, and the ensuing disturbance evolution; and (2) develop a deep reinforcement learning (DRL)-based car-following (CF) control framework with multiple objectives, including one to reduce unnecessary driver intervention. As a unique contribution, the DRL-based control is informed by two major factors: (i) evolution of the driver's distrust in automation that leads to intervention and (ii) traffic stability. 

We first model the decision-making process of voluntary intervention with an evidence accumulation model (EA) that describes the evolution of the driver distrust in AV behavior. Informed through the EA model, and human-in-the-loop driver simulations, we show how in most cases driver intervention in AVs instigates substantial traffic disturbances that are amplified along the traffic upstream.

In light of this, we propose a DRL-based CF control for AVs that systematically reduces voluntary interventions and improves traffic stability. We demonstrate the effectiveness of our model and its superiority against other controllers. Most notably, our model can reduce interventions by a margin of over 12% -30% and effectively dampen disturbances.

Presenting Author: Xinzhi Zhong University of Wisconsin-Madison

Presenting Author Biography: I am a PhD candidate at the University of Wisconsin-Madison, specializing in the development of autonomous vehicles. My research interests encompass a multidisciplinary approach, integrating machine learning, control theory, traffic flow operations, and human factors into innovative vehicular designs. Through my work, I aim to advance the field of autonomous transportation, contributing to safer, more efficient, and intelligent vehicle systems.

Authors:

Xinzhi Zhong University of Wisconsin-Madison
Yang Zhou Texas A&M University
Amudha Varshini Kamaraj University of Wisconsin-Madison
Zhenhao Zhou University of Wisconsin-Madison
Wissam Kontar University of Wisconsin-Madison
Dan Negrut University of Wisconsin-Madison
John Lee University of Wisconsin-Madison
Soyoung Ahn University of Wisconsin-Madison