Abstract
Drivers familiar with driving automation systems have been shown to exhibit improved intervention performance compared to novices, though it remains unclear whether these benefits depend on the source of the conflict. This study examined the effects of fault condition (i.e., conflict source: ego-fault vs. other-fault) and familiarity with driving automation systems on evasive performance while supervising a partial driving automation system during an impending head-on sideswipe collision. Twenty-eight licensed drivers completed four trials in a virtual reality simulator. Analyses focused on 19 participants who produced responses. Trust in automation, familiarity with driving automation systems, perception response time, ramp time, and response magnitude were measured. Results revealed a significant fault-by-familiarity interaction for steering: familiar drivers responded faster than novices in the ego-fault condition, but not in the other-fault condition. No significant effects were observed for trust, ramp time, or response magnitude. These findings suggest that familiarity effects may vary across conflicts.
Keywords
Introduction
Driver error is the dominant contributor to motor vehicle accidents, serving as the immediate pre-collision factor in most crash sequences (Singh, 2015; Treat, 1980). Many of these errors arise from limitations in human attention, perception, and decision-making. Accordingly, offloading the dynamic driving task to driving automation systems has been proposed as a potential means of improving road safety (Hester et al., 2017).
However, most motor vehicles equipped with driving automation technology available in the United States today implement “partial driving automation” (Bureau of Transportation Statistics, n.d.; Society of Automotive Engineers [SAE] International, 2021). These systems perform subtasks of the dynamic driving task and still require an in-vehicle driver to supervise the driving automation system. Thus, the driver’s role is not eliminated but shifted from driver to driving automation system supervisor.
Particularly relevant to this study are SAE Level 2 (L2) partial driving automation systems (SAE International, 2021). When engaged, these systems provide sustained lateral and longitudinal vehicle motion control, maintaining a set lane position while regulating speed and following distance. L2 system usage is limited to specified Operational Design Domains. The Operational Design Domain of a driving automation system refers to the operating conditions in which the system, or a feature of that system, is designed to function. This may include environmental (e.g., weather conditions), geographical (e.g., access-controlled freeways only), time-of-day (e.g., well-lit), and/or requisite roadway conditions (e.g., requires guard rails).
Effective human–automation teaming in these systems requires that the driver maintain situation awareness, remain vigilant, and be prepared for timely takeovers (Greenlee et al., 2018). Yet decades of research warn that this supervisory role is fragile: trust miscalibrations, vigilance decrements, reductions in situation awareness, shifts in attention, and longer takeovers are common (Eriksson & Stanton, 2017; Körber et al., 2018; Molloy & Parasuraman, 1996; Parasuraman & Riley, 1997).
Being familiar with driving automation systems may offset some of these challenges. In one study, drivers familiar with driving automation systems intervened earlier, and in a more predictive manner, than novices during a cut-in conflict (Larsson et al., 2014). Participants familiar with automated driving systems responded as soon as they detected the system accelerating in a manner suggestive of an impending conflict, whereas novices only responded once their vehicle closed on the lead vehicle to a distance they perceived as unsafe.
However, it remains unclear whether these familiarity-related advantages extend across different fault contexts, as most studies do not manipulate whether conflicts are initiated by the ego vehicle or another vehicle on the road. To address this gap, the present study examined the effects of fault condition and familiarity with driving automation systems on drivers’ evasive performance while supervising an L2 system during an impending head-on sideswipe collision.
It was hypothesized that (H1) familiar drivers would exhibit faster perception response times than novices, and (H2) this advantage for familiar drivers would be most pronounced in the ego-fault condition, yielding a fault-by-familiarity interaction. Given limited prior work on how fault condition shapes control dynamics, ramp time and response magnitude were examined. Trust in automation was measured to assess whether baseline trust was associated with performance and thus warranted inclusion as a covariate.
Method
Participants
Twenty-eight adult volunteers were recruited via word of mouth for the study. All participants were licensed to drive a motor vehicle in the United States, had normal or corrected to normal vision, and none were compensated. This study was reviewed and approved by the Institutional Review Board at California State University, Long Beach.
Virtual Environment
The virtual environment was presented using a Meta Quest 3 virtual reality headset running at 72 Hz. The device provided full six-degree-of-freedom head-tracking, allowing participants to “look around” within the vehicle (e.g., at the vehicle’s mirrors) and at the external environment. The device also provided hand-tracking, enabling participants to see their hands in the virtual environment (e.g., holding the steering wheel).
All trials took place on a flat and straight six-lane highway (three lanes in each direction; see Figure 1). The roadway did not include signs, interchanges, or entrance and exit ramps. Atmospheric conditions simulated a clear sunny day and visibility was set to 3.5 kilometers in all directions from the participant’s point of view. A highway driving environment was chosen as most documented crashes involving L2 systems occur on highways and many commercial systems are primarily designed for highway operation (Ding et al., 2024).
Fixed seated driver position.
A virtual Tesla Model 3 base model was downloaded from Sketchfab and adapted for the study (Ameer Studio, 2021). A custom C# script enabled user control of the ego vehicle via a physical steering wheel and pedal system (Simagic Alpha Drive Wheel Base, Simagic P2000 Pedal). Another custom C# script was created to simulate a driving automation system. The method ensured that all participants experienced identical route geometry, lane-change triggers, and vehicle trajectories.
Protocol
Upon arrival, participants provided informed consent, completed a pre-experiment questionnaire, and completed the Checklist for Trust between People and Automation (Jian et al., 1998).
Participants then received an overview of the hardware they would be using, the virtual environment, and the vehicle’s virtual display (see Figure 1). At this point, the ego vehicle’s driving automation system was introduced as an L2 system. Participants were informed that the driving automation system would maintain a set speed and lane position and execute lane changes when slower lead vehicles were detected. They were also told that the system could be overtaken by a braking or steering input. Finally, participants practiced manual driving, supervising the partial driving automation, and overtaking the driving automation system.
Following the training, participants completed four experimental trials. Before each trial, participants were instructed to keep their hands on the wheel, feet on the pedals, eyes on the road, and informed they could disengage the driving automation system if they perceived a hazard. The first three trials followed the same sequence for all participants. Each trial began with the ego vehicle at rest on an empty roadway. Once the participant activated the driving automation system, the ego vehicle accelerated and maintained a speed of 60 MPH while traveling in light free-flowing highway traffic (approximately six vehicles per minute traveling between 40 MPH and 80 MPH in both directions). Trials 1–3 ended with the driving automation system performing a safe lane change to bypass a lead vehicle.
The fourth trial introduced the experimental manipulation. As in the previous trials, the ego vehicle began at rest, participants activated the driving automation system, and the vehicle accelerated to a steady 60 MPH. Though in this trial, the driving automation system did not perform a safe lane change. Instead, a lateral drift was introduced that initiated the conflict. In the ego-fault condition, the ego vehicle drifted left out of its lane and encroached into the oncoming #1 lane, entering an oncoming vehicle’s path of travel (see Figure 2). In the other-fault condition, the oncoming vehicle drifted to its left over the double yellow line and encroached into the ego vehicle’s lane of travel. Fault condition order was randomized and counterbalanced across participants. Collision avoidance required an evasive maneuver by the participant. This crash configuration enabled both vehicles’ pre-crash heading angle difference, the collision type, event timing within the trial, and a one second time to collision to be held constant between conditions. No system warnings or other diagnostic cues were provided in either fault condition.
Fault condition diagrams.
Measures
Trust in automation was measured prior to the experiment using the Trust in Automation scale, a psychometrically validated instrument with demonstrated internal consistency and construct validity (Jian et al., 1998; McGrath et al., 2025). This measure was included to ensure that any group differences in response behavior were not attributable to baseline differences in trust.
Familiarity with driving automation systems was assessed prior to the experiment using a two-part self-report measure. Participants indicated whether they had any prior experience with driving automation systems (yes / no) and then provided an open-ended description of their experience. This measure was included to distinguish familiar drivers from novices, which was hypothesized to influence monitoring and response behavior during the task.
Driver performance was quantified by steering, braking, and throttle responses. In line with operationalizations of perception response time to sideswipe-threats in other studies, measurement began when the intruding vehicle changed its heading angle by more than 0.01° (Muttart, 2003). This moment is referred to as “conflict onset” throughout the remainder of the paper. Measurements ended when a response was initiated. Brake response time was defined as the interval between conflict onset and the first brake input that ultimately produced a deceleration of at least 0.2 G. Throttle response time was defined as the interval from conflict onset to the first throttle input that ultimately produced an acceleration value of 0.1 G. Steering response time was defined as the interval from conflict onset to the first steering angle input that exceeded 7°. Ramp time was measured as the interval from start of response to peak response magnitude. Response magnitude was quantified as the largest value of a response (i.e., 182 degrees of steering).
Results
Of the 28 participants, 19 produced a measurable evasive response to the conflict (e.g., braking, steering, and / or throttle input) and were retained for analysis using listwise deletion. Participants that did not respond cited a lack of time (n = 4), misallocated attention (n = 4), and one participant did not notice the conflict (n = 1). Within the subset of participants that did respond, steering was the predominant response modality; only two participants applied the brake in addition to steering, and two others applied the throttle in addition to steering. As the number of brake and throttle responses were insufficient to support reliable statistical analysis, these data were excluded for these participants, and all subsequent analyses were based on steering responses from the 19 participants (see Table 1).
Participant Demographics.
Note. M = mean. SD = standard deviation.
All statistical analyses were conducted using IBM SPSS Statistics (Version 29.0.2.0; IBM Corp, 2023). Trust was not significantly associated with steering response time, ramp time, or response magnitude, so it was not included as a covariate in the subsequent analyses. Due to space constraints, only significant results are reported.
Familiarity With Driving Automation Systems
A content analysis was conducted on participants’ descriptions of prior experience with driving automation systems. This analysis was restricted to the subset of participants who (a) produced a measurable evasive response to the conflict (n = 19) and (b) reported having prior experience with driving automation systems. This method resulted in a final sample of nine participants (n = 9). Responses reflected four primary categories of exposure.
The first category captured longitudinal driving automation features (n = 6), such as adaptive cruise control, Tesla Autopilot, non-adaptive cruise control, and cruise control with automatic braking. A second category reflected experience with lateral driving automation features (n = 5), including lane keep assist, lane centering, Tesla Full Self-Driving, and Tesla Autodrive. The third category captured active safety systems (n = 4), such as automatic emergency braking, assisted braking, and blind-spot monitoring. Finally, the fourth category involved proximity or limited exposure (n = 3), including passive experience as a passenger in automated vehicles, brief or inconsistent use of driving automation features, and general educational exposure.
The depth and type of self-reported experience with driving automation systems varied across participants; however, these differences produced a significant interaction with fault condition. Thus, familiarity was retained as a grouping factor in the subsequent analyses.
Fault-by-Familiarity Interaction
A 2 (Fault: Ego / Other) × 2 (Familiarity: Familiar / Novice) between-subjects ANOVA revealed a nonsignificant main effect of fault, F (1, 15) = 3.20, p = .094, ηp2 = .176, and a nonsignificant main effect of familiarity, F (1, 15) = 0.04, p = .845, ηp2 = .003. However, these main effects were qualified by a significant fault-by-familiarity interaction, F (1, 15) = 10.00, p = .006, ηp2 = .40 (see Figure 3).
Steering response time by fault and familiarity.
Bonferroni-corrected simple effect analyses indicated that fault impacted steering response time for the familiar group, F (1, 15) = 11.57, p = .004, ηp2 = .44, but not for the novice group, F (1, 15) = 1.00, p = .333, ηp2 = .06 (see Table 2). The nature of these effects differed across groups. The familiar group exhibited significantly faster steering response times in the ego-fault condition (M = 0.43 s) than in the other-fault condition (M = 0.98 s). Whereas the novice group exhibited nonsignificantly different steering response times across fault conditions (MEGO-FAULT = 0.80 s, MOTHER-FAULT = 0.65 s).
Performance.
Note. M = mean; SD = standard deviation.
Discussion
The results of the present study indicate that differences in steering response time emerged only among familiar drivers, who responded more quickly than novices when the ego vehicle initiated the conflict. This pattern aligns with prior work showing that drivers familiar with driving automation systems exhibit more predictive and timely interventions during system failures than novices (Larsson et al., 2014). Though the difference in steering response time between fault conditions exhibited by the familiar drivers adds nuance to this assertion. The results do not support the notion that being familiar with driving automation systems enhances supervisory performance across all conflict contexts. Instead, the advantage was specific to conflicts initiated by the ego vehicle. It is possible that familiar drivers allocated more monitoring resources to supervising the ego vehicle and its driving automation system at the cost of monitoring the broader driving environment. Future research incorporating eye-tracking or other attention-monitoring methods may clarify the cognitive processes underlying these differences.
While not statistically significant, the ramp time and response magnitude patterns offer insight into drivers’ evasive performance. When the ego vehicle initiated the hazard, familiar drivers produced larger steering magnitudes and longer ramp times. This pattern may reflect the additional time required to reach those higher peaks. Although in that same condition, the novices produced comparatively small steering magnitudes while still requiring substantial time to reach them. These measures suggest that when the ego vehicle initiated the conflict, the two groups adopted different evasive strategies. Familiar drivers responded faster with a larger corrective input than novices, who responded more slowly and with a smaller magnitude. When the other vehicle initiated the conflict, both groups produced similar ramp times and magnitudes, reflecting a more uniform evasive strategy.
Limitations
Virtual reality simulations are used in driving research to increase experimental control while preserving key perceptual and behavioral features of real-world driving; however, they do not fully reproduce the sensory, cognitive, and motor demands of operating a vehicle. The experimental trial lasted 7 min, but driving automation system monitoring can extend for much longer durations. Extended monitoring is known to produce vigilance decrements, which can impair drivers’ ability to respond to conflicts. Real-world performance may be poorer than what was observed in this study. The simulated driving environment was simplified relative to typical roadway conditions. The roadway did not include signs, interchanges, heavy traffic, or entrance and exit ramps. These elements may increase task demands and leave fewer attentional resources for monitoring the roadway and driving automation system. This study used a brief self-report measure for distinguishing between familiar and novice drivers. More detailed usage metrics may refine how different depths and types of familiarity shape performance. A small, non-random convenience sample was employed for this study. Although 28 participants completed the experiment, only 19 responded and were retained for analyses. A post hoc analysis indicated that the study was underpowered for detecting small- and medium-sized effects.
Conclusion
These findings refine existing assumptions about the advantages exhibited by drivers familiar with driving automation systems, suggesting that familiar drivers may not demonstrate enhanced performance across all conflict contexts. As partial driving automation becomes more widespread, understanding how drivers of different familiarity levels respond to different conflicts while supervising these systems will be critical for improving roadway safety.
Footnotes
Acknowledgements
The authors thank Anushka Brito for her assistance with data post-processing and analysis. The authors also thank the Master of Science in Human Factors graduate program at California State University, Long Beach for its support and collaboration on this research. Aperture, LLC provided the research facility, personnel time, and resources to support this research; however, the analyses and conclusions reflect the independent scholarly work of the authors and do not represent the official positions or views of the company.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and / or publication of this article.