Physiological Evaluation of User Experience in Unstable Automated Driving: A Comparative Study

Abstract

While automated-driving technology is advancing rapidly, human-centered research is still in its early stages. Research on negative user responses to automated driving is particularly limited in complex roadway environments such as roundabouts, where driving decisions typically depend on driver judgment and traffic conditions. In these environments, automated-driving vehicles may exhibit unstable behaviors, such as sudden stops or forced intersection entries. Since successful interaction between users and automated systems is critical for widespread adoption, understanding when and how automated driving negatively affects users is essential. This study investigated user psychological responses and corresponding physiological changes during unstable automated-driving situations. Using a virtual environment driving simulator, we compared two scenarios: sensor-only automated driving (A.D(S)), which exhibited unstable driving patterns; and cooperative automated driving (A.D(C)), which achieved more stable performance through infrastructure communication. We analyzed the responses of 30 participants using electromyography (EMG) measurements and pupil diameter tracking, supplemented by qualitative evaluations. Results showed that A.D(S) participants experienced higher levels of frustration during prolonged waiting times compared to A.D(C) participants. In addition, sudden braking events elicited startle responses characterized by pupil dilation and elevated arm-muscle EMG readings. This research advances our understanding of how automated-driving behaviors affect user experience and emphasizes the importance of human factors in the development of automated-driving technologies.

1. Introduction

With the rapid advancement of automated-driving technology and ongoing demonstration projects, public expectations for automated-driving vehicles continue to rise [1,2]. Globally, automobile manufacturers and IT companies are making significant investments in automated-driving technologies, while countries establish institutional frameworks for their commercialization [3]. Worldwide, many participants have eagerly tested these vehicles, experiencing this technological innovation firsthand while anticipating its broader commercial deployment.

While research has primarily focused on developing automated-driving capabilities, these technologies must evolve to provide comfortable driving experiences that users can trust and rely on [4,5,6,7]. The goal of automated-driving services extends beyond technical perfection to delivering safe and comfortable transportation. User acceptance requires technology development that integrates personal factors with system characteristics [4,6], alongside user experience (UX) research focused on user comfort [8,9]. Currently, automated-driving technology has not yet reached a level where users feel consistently safe—a major barrier to widespread adoption of these services [5,10,11].

Automated-driving vehicles face limitations in maintaining stable driving performance through technology alone. While they can achieve smooth operation on continuous roads and signalized intersections using onboard sensors (LiDAR, radar, and cameras) and HD map technology [12,13,14], as these environments follow predictable, standardized traffic rules. However, they struggle to maintain stable performance in more complex environments, like roundabouts and unsignalized intersections, where navigation based solely on automated technology proves challenging [15,16,17].

In roundabouts, automated-driving vehicles may experience excessive latency when determining distances between entering vehicles or attempt dangerous entries that risk collisions with vehicles already in the circular lane [15]. This occurs because automated algorithms cannot fully replicate human drivers’ intuitive judgment and communication skills. Similar unstable behavior emerges at unsignalized intersections and junctions, where vehicles must assess inter-vehicle distances and make entry decisions based on traffic flow. Further complications arise when objects suddenly emerge from sensor blind spots, triggering emergency braking or risking collisions. These situations can lead to unstable driving behaviors, such as missed entry opportunities or sudden braking, resulting in passenger frustration and startle responses [18,19,20,21,22,23].

Unstable automated driving can trigger negative psychological changes, like frustration and shock, leading to measurable physiological responses [24,25,26,27,28,29,30]. During stable automated driving, users show decreased heart rate and reduced visual load [24]. However, stress and anxiety can manifest in increased electromyography (EMG) [25,27], elevated heart rate [26], and pupil dilation [27]. These unpleasant experiences can erode confidence in and acceptance of automated driving [31]. Thus, despite technological advances, failure to ensure user safety and comfort may fundamentally hinder widespread adoption.

Physiological measurements provide quantitative data about how emotions in specific situations affect the body. Electromyography (EMG), which records muscle contraction signals and resting currents, has applications across kinesiology, rehabilitation medicine, and human engineering [32,33,34,35]. In driving research, EMG serves as an indicator of physiological changes [36,37,38,39], with values increasing when drivers experience negative emotions or higher mental workload [37,38,39]. Negative emotions, such as frustration or startle responses, alter motor neuron activation patterns, producing measurable changes in muscle activity that EMG can detect [34,35,37,38].

Changes in pupil diameter represent a key physiological response to startle and have been extensively studied. Pupil dilation occurs due to increased sympathetic nervous system activity and intensifies in response to emotional stimuli or stress [40,41,42]. Researchers have utilized these pupillary changes as physiological indicators to measure drivers’ cognitive load and evaluate their level of startle response in various situations [43,44,45].

In this context, the importance of technological development that considers the user experience in addition to the technical performance of automated-driving vehicles is becoming increasingly important [46,47,48]. Ride comfort, driving stability, and user interaction are key elements for the successful introduction and diffusion of automated-driving services. In particular, the response technology in traffic obstruction situations and securing user trust will be critical challenges for the future development of automated-driving technology.

Meinlschmidt et al. (2019) emphasized the need for systematic research on users’ psychobiological responses to automated driving. Studies in simulated environments have shown increased heart rate and electromyography responses during automated-driving scenarios [49]. He et al. (2022) analyzed changes in users’ risk perception and confidence during sudden braking and lane changes through simulator experiments. Using pre- and post-experiment surveys and physiological measurements like heart rate and pupil diameter, they found that higher perceived risk correlated with lower confidence and increased physiological responses, such as pupil dilation [50]. Similarly, Beggiato et al. (2019) evaluated physiological responses to six potentially uncomfortable traffic situations in automated driving, including intersection hazards, evasive maneuvers, and highway entry [29]. Their findings revealed that more unexpected situations triggered greater physiological changes.

While early research relied primarily on simulations, recent studies have expanded to real automated-driving vehicles and real-world ride quality evaluation. One study evaluated passenger comfort and anxiety through physiological signals during actual automated driving [30]. The research implemented various driving events, such as overtaking, stopping, and sharp turns, finding that participant stress levels peaked during stopped states. In-vehicle evaluation provides valuable insights, as users can fully experience and evaluate automated-driving behaviors. However, these real-world studies face limitations: they are restricted to specific road environments for safety, require experimenter presence, and cannot control weather conditions.

This study experimentally verifies the importance of user experience in automated driving-technology development. Specifically, we analyzed users’ psychological responses when automated driving deviates from expectations, using both qualitative evaluation and physiological indicators (EMG and pupil diameter). We defined potential traffic interference situations during unstable automated driving and measured participants’ psychological changes using a driving simulator in a virtual road environment. This study’s significance lies in its comprehensive analysis of how abnormal automated driving-vehicle behavior affects user experience, combining qualitative evaluation with physiological response measurements.

2. Research Methodology

In this study, experimental research was conducted using a virtual environment-based driving simulator to evaluate the negative impact on users in situations where automated driving is impeded. Using simulation software, the study implemented road environments and designed sections, such as roundabouts, where automated-driving constraints could occur, simulating situations that could affect users. The research evaluated participants’ experiential observations of automated-driving situations through survey-based qualitative evaluation. The study analyzed EMG and pupil diameter measurements to verify that the effects felt by the participants were reflected in actual physiological changes.

2.1. Experiments with a VR Simulator

2.1.1. Apparatus

This study utilized a specialized driving simulator with a semi-dome display at the University of Seoul VR Center. The simulator features authentic vehicle components—including a steering wheel, accelerator and brake pedals, instrument panel, gear shift, and tilt lever—providing an authentic vehicle cockpit experience (Figure 1). The integrated semi-dome display enhances participant immersion in the virtual environment, enabling realistic driving conditions. This simulator has proven effective for user evaluation studies by facilitating both natural automated-driving movements and participant-response measurements in a virtual environment [51,52].

The VR environment and driving scenarios for the experiment were created using UC-win/Road (ver. 16.0), developed by FORUM8. UC-win/Road is widely used in simulator experiments due to its easy road geometry creation and traffic environment-editing capabilities. It also provides various vehicle behaviors’ configuration features. Thus, studies that have implemented a virtual road environment and conducted experiments in conjunction with a simulator have shown that the results are similar to those of experiments conducted on real roads [53]. Thus, software programs such as UC-winRoad (ver.15.1.4) that are linked to simulators are used in a variety of ways to replicate situations that are difficult to experiment with in an actual road environment in terms of user behaviors [51,54,55].

To evaluate the actual level of frustration or startle experienced by the participants during the experiment, EMG measurements were used. EMG measurement was performed using the NORAXON Telemyo DTS system, as shown in Figure 2a. This device enhances experimental efficiency by wirelessly detecting and recording electrical signals generated during muscle activity. It is compatible with EMG and other biomechanical sensors. The device can wirelessly collect data from up to 10 m away, making it suitable for experimental research [56]. In this study, sensors were attached to the right arm at the biceps and to the right leg at the calf muscle to measure EMG activity, as shown in Figure 2c. To measure the EMG values with this device, the bandwidth was set to 20–400 Hz, and the sampling rate was set to 1500 Hz to measure the participant’s EMG signals that occur during the experiment.

Pupil diameter changes, another physiological indicator, were measured using Tobii Pro Glasses 3 (Figure 2b). This spectacle-format device measures visual behaviors in real time at millisecond intervals and is comfortable to wear. It also accommodates people with low vision through optional lens accessories.

2.1.2. Experimental VR Scenario Design

To evaluate the negative effects experienced by users during unstable automated driving, different road environments were implemented through virtual-environment simulations. As shown in Figure 3, the course was designed to alternate between urban roads with two one-way lanes and highways with three one-way lanes, allowing users to naturally experience typical automated driving in road environments. To simulate situations where users might experience frustration, the course included roundabouts, highway merge sections, and unsignalized intersections where gap acceptance could be delayed. In addition, to create situations where users might experience startle, the course included unsignalized intersections and crosswalk sections where pedestrians suddenly appear from blind spots. To avoid interference between the effects of different sections, major sections were separated by distances of more than 500 m, and the design alternated between normal driving sections and sections where specific events occurred.

2.1.3. Automated-Driving Implementation

Two scenarios were implemented to evaluate how unstable automated driving affects users. The first scenario (A.D(S)) uses automated-driving technology alone, while the second scenario (A.D(C)) employs cooperative automated driving for stable navigation in traffic-obstructed sections. Both scenarios operate identically under normal conditions but show different behaviors when encountering obstacles like roundabouts, merging sections, unsignalized intersections, and blind-spot pedestrian crossings.

As shown in

Table 1. Setting values of driving characteristics based on automated-driving styles.

Categories		A.D(S)	A.D(C)
Driving speed (km/h)	Urban road	60 km/h
	Highway	80 km/h
Acceleration ($\mathrm{m}/{\mathrm{s}}^2$)		1.0 $\mathrm{m}/{\mathrm{s}}^2$
Deceleration ($\mathrm{m}/{\mathrm{s}}^2$)		−2.8 $\mathrm{m}/{\mathrm{s}}^2$
Gap acceptance (s)		5 s ≤	4 s ≤
Etc.		Driving assumption based on traffic evaluation using only vehicle sensors and systems	Driving assumption based on traffic evaluation while sharing information with surrounding vehicles and infrastructure

, considering the characteristics of the Korean road environment, the driving speeds for both scenarios were set at 60 km/h for urban roads and 80 km/h for highways. The acceleration and deceleration values were applied based on the automated-driving characteristics preferred by the majority in an automated driving-preference evaluation study [51].

The key distinction between scenarios lies in gap acceptance at roundabouts, highway merges, and unsignalized intersections. A.D(S) requires gaps of 5 s or more, while A.D(C) requires 4 s or more. While gap acceptance standards for automated driving are not yet clearly defined, the Highway Capacity Manual (HCM) suggests a basic critical gap of 5.19 s [57], and some studies propose a minimum of 3.6 s [58]. Currently, automated-driving vehicles being tested in Korea typically wait 5 s when encountering traffic-flow obstacles [59].

For this study, A.D(S) uses a 5 s gap acceptance time, reflecting conservative behavior based on existing practices [57,59]. A.D(C) uses 4 s, considering its cooperative capabilities and research on critical gaps in advanced automated driving [57,58]. These thresholds balance operational efficiency with safety—the longer 5 s gap in A.D(S) compensates for sensor-only operation limitations, while A.D(C)’s 4 s gap reflects enhanced awareness from cooperative systems while maintaining safety. This difference allows for a comparison of how each scenario’s behavior affects users.

2.1.4. Implementation of Events for Automated-Driving Evaluation

As shown in

Table 2. Detailed implementation of event situations by scenario.

Categories	A.D(S)	A.D(C)
Roundabout	- Post-stop departure : Gap acceptance after 8 rejections of gaps under 5 s	- Post-stop departure : Gap acceptance after 4 rejections of gaps under 4 s
Merging	- Post-stop departure : Gap acceptance after 8 rejections of gaps under 5 s	- Low-speed merge sequence : Gap acceptance after 3 rejections of gaps under 4 s
Unsignalized intersection	- Sudden stop activation due to blind-spot latency in oncoming vehicle detection - Post-stop departure : Gap acceptance after 7 rejections of gaps under 5 s	- Controlled deceleration to stop based on early detection of oncoming vehicle - Post-stop departure : Gap acceptance after 3 rejections of gaps under 4 s
Crosswalk	- Sudden stop activation and subsequent departure due to blind-spot latency in pedestrian detection	- Controlled stop and departure based on early pedestrian detection

, two test scenarios were implemented to evaluate users’ frustration and startle responses during automated driving: (1) gap acceptance in traffic-heavy sections (roundabouts, highway merging, and unsignalized intersections) and (2) sudden braking responses to unexpected obstacles (opposing vehicles and jaywalking pedestrians at blind-spot intersections).

For frustration evaluation, the study simulated waiting situations due to continuous traffic flow at roundabouts, highway merges, and unsignalized intersections. Both scenarios maintained vehicle gaps of 2.5–6 s, but A.D(S) was set to reject approximately twice as many gaps as A.D(C). Specifically, when a 4-s or larger gap first appeared, A.D(C) would accept it while A.D(S) would reject it, allowing measurement of any resulting increase in user frustration.

For startle-response evaluation, the study tested sudden stops in blind spots where self-driving sensors alone struggle to detect objects. A.D(S) implemented sudden stops due to conflict vehicles at unsignalized intersections and unexpected pedestrians. In contrast, A.D(C) simulated receiving advance information from surrounding infrastructure, allowing for gradual deceleration at the same points.

2.1.5. Automated-Driving Evaluation

The evaluation of A.D(S) and A.D(C) scenarios combined qualitative questionnaires with physiological measurements (EMG and pupil diameter). The qualitative assessment used van der Laan’s Acceptance Scale [60], which enhances reliability by having participants rate their emotional responses across nine contrasting emotions to evaluate overall usefulness and satisfaction.

The standard van der Laan questionnaire was applied to both the complete driving experience and specific sections, like roundabouts and merging areas. This aimed to compare satisfaction levels between stable cooperative driving and unstable automated driving. Additionally, “frustration due to waiting time” and “degree of surprise” were rated separately on a five-point scale post-experiment, as these metrics were not included in van der Laan’s scale.

For physiological analysis, participants’ EMG readings from the right arm (biceps) and leg (calf muscle) were recorded to correlate with frustration and startle responses. Pupil diameter measurements were also used to analyze correlations with startle-response levels.

2.2. Participants

A total of 32 participants took part in the study. It is desirable to obtain as many samples as possible to test the differences in physiological changes on the user side that occur in the unstable automated-driving situations defined in this study [61,62], but it is difficult to conduct large-scale experiments on humans, and similar previous studies have also recruited about 30 subjects to conduct experiments to obtain meaningful results [28,29,30]. And considering that the purpose of this study is to identify trends and suggest the need for related research, we plan to conduct the experiment on at least 30 subjects. In the recruitment process, we recruited a relatively small proportion of participants over the age of 50 who may experience difficulties such as motion sickness in a driving simulator environment, and participants in their 20s who have relatively little actual driving experience. We reduced the proportion of participants over the age of 50 to prepare for situations in which motion sickness, which can occur in a simulator environment, may affect the health of participants or make it difficult to conduct normal experiments [63,64]. In fact, of the 32 participants, 1 woman in her 50s and 1 woman in her 40s experienced motion sickness and complained of dizziness and vomiting during the experiment, and the experiment was discontinued. Also, for the purpose of this experiment, the number of participants in their 20s was reduced in order to recruit people with relatively more than a few years of driving experience, as the participants were to judge the driving characteristics that occur in unstable automated-driving situations based on their own driving experience. For the final analysis, data from 30 participants were used, as shown in

Table 3. Status of participants used in the analysis.

Ages	Male	Female	Total
20s	2	2	4
30s	8	3	11
40s	8	3	11
50s	3	1	4
Total	21	9	30

. Looking at the driving experience of the participants whose data were used, we can see that the distribution ranges from a minimum of 3 years to a maximum of 30 years, as shown in Figure 4.

2.3. Experiment Procedure

As shown in Figure 5, the participant in the drvier’s seat was fitted with the glasses device to measure the diameter of the pupil and was attached to an electromyograph on the biceps of the right arm and the calf muscles of the right leg. Both pupil diameter and EMG were measured using the dedicated program installed on a separate computer, and the EMG measured in the two areas was recorded simultaneously using the dedicated program.

The time of the dedicated program and the simulator equipment were synchronized, and two experimenters simultaneously ran each dedicated program on their own computers to check whether each physiological datum was synchronized properly. After each experiment, the time when each piece of biometric information was measured was compared with the time when the experiment scenario was performed to ensure that the data were measured correctly.

To eliminate any learning effect, the order of the scenarios was randomized, and the survey was administered immediately after each scenario so that participants could immediately evaluate their experience.

2.4. Analysis Method

To analyze the user’s evaluations of the A.D(S) and A.D(C) scenarios, we comprehensively used the participants’ qualitative rating results and measured physiological changes. For the qualitative rating results, we directly used the 5-point scale ratings from the participants’ responses. The section-specific qualitative ratings for frustration and startle were linked to measured physiological changes to increase the credibility of the qualitative ratings.

Among the measured physiological changes, the EMG data required normalization due to large differences in measurements between participants, and a number of preprocessing steps were performed to achieve this. Existing studies dealing with EMG data have also performed normalization to correct for measurement differences between participants [36,37,38,65,66]. Two main normalization methods were used. Maximum Voluntary Contraction (MVC) is a method that measures the subject’s maximum EMG value and normalizes the EMG value based on that value [65]. This is mainly used in studies where the measurement is based on the actual activity of the subject. Another method is the min–max normalization method, which uses the minimum and maximum values of the recorded EMG values for each individual subject [66]. This is used to analyze EMG data recorded in an experimental setting, where it is difficult to separately measure the maximum active value of the EMG data, or where the subject is not required to move much. In this study, the min–max normalization method was used because the EMG data were analyzed in a situation where the participants’ movements were not required.

In this study, the following preprocessing steps were taken to use the measured EMG data for analysis. First, outliers recorded at abnormally high levels were removed. The outliers were removed using the method of adding or subtracting 3 standard deviations from the mean [36,37,67]. Data outside the range of 3 standard deviations from the mean were defined as outliers and removed. After outlier removal, the Kalman filter method was applied to smooth the time-series EMG data to maintain the continuity of the removed outlier values and to compensate for measurement delays [68]. Then, the 1500 EMG values recorded per second were grouped into 150 groups in sequence, and the root mean square value for each group was calculated to derive the EMG values at 0.1 s intervals. Based on the preprocessed data, minimum–maximum normalization was performed to organize the EMG measurement results of each participant for analysis.

$${\chi }^{\prime }={\displaystyle \frac{\chi -\mathrm{m}\mathrm{i}\mathrm{n}\left(\chi \right)}{\max \left(\chi \right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(\chi \right)}}$$

(1)

where ${\chi }^{\prime }$ is the normalized value, $\chi$ is the original value, min($\chi$) is the minimum value in the dataset, and max($\chi$) is the maximum value in the dataset.

For pupil diameter, measurement results obtained by the device’s dedicated analysis program, which records pupil diameter at 1/1000 s intervals, were exported as Excel data and used for analysis.

As shown in

Table 4. Temporal segmentation of discrete intervals for physiological change analysis.

Roundabout			Merging			Unsignalized Intersection			Crosswalk
Seg.	A.V(S)	A.V(C)	Seg.	A.V(S)	A.V(C)	Seg.	A.V(S)	A.V(C)	Seg.	A.V(S)	A.V(C)
R-1	Pre-entry (10 s)		M-1	Pre-entry (10 s)		U-1	Pre-entry (10 s)		C-1	Pre-entry (10 s)
R-2	Stop at stop line ($\pm$3 s)		M-2	Wait (15 s)	Slow-driving (5 s)	U-2	Stop at stop line ($\pm$3 s)		C-2	Sudden stop activation ($\pm$3 s)
R-3	Gap rejection (10 s)		M-3	Merging (5 s)		U-3	Gap rejection (10 s)
R-4	Rejection of 4 s gap (6 s)	Gap acceptance (6 s)	M-4	Driving on highway (10 s)		U-4	Rejection of 4 s gap (6 s)	Gap acceptance (6 s)
R-5	Gap rejection (10 s)					U-5	Gap acceptance (6 s)
R-6	Gap acceptance (6 s)

, average values of EMG (arm and leg) and pupil diameter were calculated for specific moments to analyze in detail the physiological changes exhibited by the participants for each target section. For the roundabouts, the analysis was divided into 6 moments for the A.D(S) scenario and 4 moments for the A.D(C) scenario. Using the moment before entering the roundabout (Seg. R-1) as the baseline, paired t-tests were performed to verify differences in EMG and pupil diameter measurements compared to the 3 s moment of stopping at the stop line after entering (Seg. R-2), moments of continuous short gap rejection (Seg. R-3), the moment of rejection (A.V(S)) or acceptance (A.V(C)) of a 4 s gap (Seg. R-4), subsequent moments of short gap rejection (Seg. R-5), and finally the moment of gap acceptance in the A.V(S) scenario (Seg. R-6).

Similar segmentation was applied to highway merge sections and unsignalized intersections. For highway merge sections, the analysis compared differences from the moment of driving on the ramp before entering the acceleration lane (Seg. M-1) with the moments of waiting (A.D(S)) or driving slowly (A.D(C)) in the acceleration lane (Seg. M-2), the moment of merging onto the mainline (Seg. M-3), and the moment of driving on the highway after merging (Seg. M-4). For unsignalized intersection sections, the analysis compared the moment before entering the intersection (Seg. U-1) with the moment of stopping at the stop line (Seg. U-2), the moments of continuous short gap rejections (Seg. U-3), the moment of rejecting (A.D(S)) or accepting (A.D(C)) a 4 s gap (Seg. U-4), and finally the moment of gap acceptance in the A.D(S) scenario (Seg. U-5). For sudden-stop situations due to pedestrians appearing in blind spots, the analysis compared the moment before the event occurred (Seg. C-1) with the moment of sudden stop when the event occurred (Seg. C-2).

2.5. Statistical Analysis Method

In this study, a paired t-test was used to determine whether the differences in the measured physiological changes were statistically significant. The paired t-test is a statistical analysis method that tests whether the differences in the measured values at two different times for the same subject are statistically significant, and it is a test that compares the means of two dependent groups that are related to each other. In addition, a Cohen’s d value was derived to analyze the effect size of the differences between the two dependent groups [69]. Cohen’s d is an effect size index that standardizes the average difference between two groups. The d value is calculated as a value between 0 and 1, and the higher the value, the larger the effect size. A value between 0.1 and 0.3 is considered a small effect size, a value between 0.4 and 0.7 is considered a moderate effect size, and a value between 0.8 and 1 is considered a large effect size. In this study, the differences in physiological changes in each section defined for each road environment were tested using a paired t-test, as shown in Table 4. The statistical analysis results were obtained using the SPSS program (ver 20.0).

3. Results

3.1. User Evaluation of the Overall Automated-Driving Style

Regarding the participants’ ratings of usefulness and satisfaction for the overall driving experience, A.D(C) received more positive ratings than A.D(S), as shown in

Table 5. Evaluation results of usefulness and satisfaction by A.D style (n = 30, 5 scale).

Categories		A.D(S)	A.D(C)
Overall driving	Usefulness	3.12	3.79
	Satisfaction	2.71	3.93
Roundabout	Usefulness	2.63	3.61
	Satisfaction	2.26	3.73
Merging	Usefulness	2.66	3.03
	Satisfaction	2.24	2.89
Unsignalized intersection	Usefulness	2.37	3.63
	Satisfaction	2.15	3.65
Crosswalk	Usefulness	1.92	3.44
	Satisfaction	2.05	3.52

. A.D(C) scored 0.67 points higher in usefulness and 1.22 points higher in satisfaction than A.D(S), indicating that a stable driving style in automated driving results in higher satisfaction. Notably, A.D(C) received relatively higher ratings for roundabouts, unsignalized intersections, and pedestrian crossings.

While A.D(S) generally received low usefulness and satisfaction scores, there were differences in the evaluation results depending on the section’s characteristics. At roundabouts and highway merge sections, where frustration was induced, usefulness was rated relatively higher compared to satisfaction. However, at unsignalized intersections and pedestrian crossings, where startle was induced, both usefulness and satisfaction were rated low. This suggests that startle has a more negative effect than frustration on the evaluation of automated driving.

3.2. Evaluation of User Response Under Extended Delay Conditions

When rating frustration due to waiting in automated-driving situations at roundabouts, highway merges, and unsignalized intersections, participants experienced high levels of frustration, as confirmed by significant increases in EMG (leg) measurements. As shown in

Table 6. Results of user frustration ratings and physiological change analysis at roundabout segmentations (n = 30).

Categories		A.D(S)				A.D(C)
		Mean	S.D.	Mean Difference (ρ-Value) *	Cohen’s d	Mean	S.D.	Mean Difference (ρ-Value) *	Cohen’s d
Self-rating of frustration (5 scale)		3.84	-	-	-	1.82	-	-	-
EMG (arm) (0–1 n.u.)	Seg. R-1	0.351	0.212	-	-	0.333	0.142	-	-
	Seg. R-2	0.360	0.204	0.009 (0.800)	0.048	0.341	0.174	0.007 (0.857)	0.034
	Seg. R-3	0.339	0.213	−0.012 (0.796)	0.048	0.355	0.150	0.022 (0.516)	0.122
	Seg. R-4	0.336	0.202	−0.016 (0.632)	0.090	0.338	0.151	0.004 (0.897)	0.024
	Seg. R-5	0.345	0.206	−0.006 (0.887)	0.027	-	-	-	-
	Seg. R-6	0.345	0.187	−0.006 (0.901)	0.023	-	-	-	-
EMG (leg) (0–1 n.u.)	Seg. R-1	0.350	0.180	-	-	0.345	0.185	-	-
	Seg. R-2	0.366	0.203	0.016 (0.668)	0.080	0.373	0.173	0.027 (0.441)	0.145
	Seg. R-3	0.387	0.223	0.037 (0.421)	0.152	0.381	0.147	0.036 (0.257)	0.215
	Seg. R-4	0.424	0.228	0.074 (0.042) **	0.396	0.370	0.182	0.025 (0.485)	0.131
	Seg. R-5	0.386	0.227	0.036 (0.373)	0.168	-	-	-	-
	Seg. R-6	0.377	0.208	0.027 (0.562)	0.109	-	-	-	-

* Numerical difference between the current segmentation and the first segmentation. ** ρ < 0.05.

, the frustration experienced by participants at roundabouts was rated at 3.84 points for A.D(S) and 1.82 points for A.D(C), with A.D(S) showing approximately twice the level of frustration. These subjective ratings were verified objectively by changes in EMG (leg) measurements. Compared to the moment before entering the roundabout (Seg. R-1), EMG (leg) measurements generally increased in subsequent segments, with the most significant increase occurring in the A.D(S) scenario when rejecting the 4 s gap at Seg. R-4. This showed a statistically significant average increase of 0.074 compared to the pre-entry segment. The Cohen’s d value for the corresponding value is approximately 0.4, which confirms that the effect size is moderate. In contrast, in the A.D(C) scenario, EMG (leg) measurements showed a slight increase up to Seg. R-3 and then showed a decreasing trend at Seg. R-4 as gap acceptance occurred. The instantaneous-change pattern of the EMG (leg) measurement at the roundabout is shown in Figure 6, and the distribution of the EMG (leg) values measured by segment is shown in Figure 7.

Table 7. Results of user frustration ratings and physiological change analysis at merging segmentations (n = 30).

Categories		A.D(S)				A.D(C)
		Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d	Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d
Self-rating of frustration (5 scale)		3.39		-	-	2.37		-	-
EMG (arm) (0–1 n.u.)	Seg. M-1	0.377	0.200	-	-	0.389	0.167	-	-
	Seg. M-2	0.449	0.238	0.072 (0.022) **	0.452	0.359	0.177	−0.031 (0.377)	0.167
	Seg. M-3	0.379	0.214	0.002 (0.969)	0.007	0.375	0.168	−0.014 (0.737)	0.063
	Seg. M-4	0.343	0.196	−0.034 (0.426)	0.150	0.357	0.166	−0.033 (0.412)	0.155
EMG (leg) (0–1 n.u.)	Seg. M-1	0.378	0.173	-	-	0.390	0.162	-	-
	Seg. M-2	0.408	0.176	0.029 (0.394)	0.161	0.373	0.152	−0.017 (0.556)	0.111
	Seg. M-3	0.377	0.203	−0.001 (0.983)	0.004	0.357	0.189	−0.033 (0.398)	0.159
	Seg. M-4	0.356	0.199	−0.023 (0.632)	0.090	0.364	0.149	−0.026 (0.340)	0.180

* Numerical difference between the current segmentation and the first segmentation. ** ρ < 0.05.

shows participant ratings for A.D(S) and A.D(C) in highway merge situations. Participants reported relatively high frustration levels of 3.39 points for A.D(S), and even for A.D(C), they reported frustration levels of 2.37 points, a measurement which was higher compared to that of other segments. Analysis of sequential physiological changes revealed significant changes in EMG (arm) measures at Seg. M-2, which measures driving behaviors in the acceleration lane, as shown in Figure 8 and Figure 9a. In particular, in the case of A.D(S), the EMG (arm) measurement value increased by 0.072 compared to segment M-1 due to dissatisfaction and frustration with the automated-driving vehicle stopping instead of accelerating when merging, which was a statistically significant result, and the Cohen’s d value was also approximately 0.45, confirming that the effect size was moderate. In contrast, in A.D(C), EMG measurements at Seg. M-2 showed slight decreases in both arm and leg measurements compared to Seg. M-1. This suggests that the continuous driving behaviors of controlling speed by driving slowly to ensure adequate gaps in the acceleration lane helped to reduce participant frustration.

As shown in

Table 8. Results of user frustration ratings and physiological change analysis at non-signalized intersections (n = 30).

Categories		A.D(S)				A.D(C)
		Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d	Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d
Self-rating of frustration (5 scale)		3.06	-	-	-	1.56	-	-	-
EMG (arm) (0–1 n.u.)	Seg. U-1	0.333	0.187	-	-	0.343	0.172	-	-
	Seg. U-2	0.378	0.219	0.044 (0.232)	0.227	0.322	0.193	−0.021 (0.620)	0.034
	Seg. U-3	0.343	0.176	0.010 (0.780)	0.052	0.317	0.162	−0.026 (0.543)	0.122
	Seg. U-4	0.352	0.233	0.019 (0.669)	0.080	0.315	0.202	−0.028 (0.587)	0.024
	Seg. U-5	0.348	0.213	0.015 (0.741)	0.062	-	-	-	-
EMG (leg) (0–1 n.u.)	Seg. U-1	0.339	0.129	-	-	0.393	0.206	-	-
	Seg. U-2	0.402	0.192	0.062 (0.125)	0.294	0.364	0.169	−0.029 (0.442)	0.145
	Seg. U-3	0.435	0.176	0.095 (0.006) **	0.547	0.343	0.166	−0.050 (0.203)	0.242
	Seg. U-4	0.424	0.161	0.085 (0.018) **	0.468	0.330	0.189	−0.063 (0.103)	0.313
	Seg. U-5	0.410	0.209	0.071 (0.055)	0.372	-	-	-	-

* Numerical difference between the current segmentation and the first segmentation. ** ρ < 0.05.

, the evaluation results for the unsignalized intersection situations revealed that participants reported frustration levels of 3.06 points for A.D(S) and 1.56 points for A.D(C), showing similar trends to the other segments. Significant physiological changes in this segmentation were observed primarily in the EMG (leg) measures. As shown in Figure 10 and Figure 11, in the A.D(S) scenario, significant measurement differences began to occur from the moment of stopping at the stop line (Seg. U-2) compared to the moment before entering the unsignalized intersection (Seg. U-1). Specifically, the EMG (leg) measurements showed a maximum increase of 0.095 at Seg. U-3 compared to Seg. U-1, which is interpreted as the frustration of waiting affecting the EMG (leg) measurement changes. The Cohen’s d value for the average difference is approximately 0.55, which indicates that a relatively large effect size has been derived. In contrast, the A.D(C) scenario showed a general decreasing trend in EMG (arm and leg) measurements in the following segments compared to Seg. U-1. These results suggest that the A.D(C) driving style provided the participants with a sense of stability and satisfaction.

3.3. Evaluation of User Response to Unexpected Events

When evaluating the startle response to sudden situations in automated driving at unsignalized intersections and pedestrian crossings, participants experienced a startle response during the sudden stops of automated driving, which was confirmed by significant increases in EMG (arm) measurements and pupil diameter. As shown in

Table 9. Results of user startle response ratings and physiological change analysis at unsignalized intersections (n = 30).

Categories		A.D(S)				A.D(C)
		Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d	Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d
Self-rating of startle response (5 scale)		2.56		-		1.28		-	-
EMG (arm) (0–1 n.u.)	Seg. U-1	0.333	0.187	-	-	0.343	0.172	-	-
	Seg. U-2	0.378	0.219	0.044 (0.232)	0.227	0.322	0.193	−0.021 (0.620)	0.034
	Seg. U-3	0.343	0.176	0.010 (0.780)	0.052	0.317	0.162	−0.026 (0.543)	0.122
	Seg. U-4	0.352	0.233	0.019 (0.669)	0.080	0.315	0.202	−0.028 (0.587)	0.024
	Seg. U-5	0.348	0.213	0.015 (0.741)	0.062	-	-	-	-
Pupil diameter (mm)	Seg. U-1	3.901	0.663	-	-	3.949	0.609	-	-
	Seg. U-2	3.994	0.621	0.093 (0.042) **	0.395	3.834	0.633	−0.115 (0.002) **	0.634
	Seg. U-3	3.934	0.682	0.034 (0.449)	0.143	3.855	0.572	−0.094 (0.060)	0.364
	Seg. U-4	3.861	0.596	−0.040 (0.789)	0.050	3.741	0.573	−0.209 (0.001) **	0.916
	Seg. U-5	3.828	0.615	−0.072 (0.085)	0.332	-	-	-	-

* Numerical difference between the current segmentation and the first segmentation. ** ρ < 0.05.

, startle-response levels at unsignalized intersections were rated at 2.56 points for A.D(S) and 1.28 points for A.D(C), with A.D(S) showing relatively higher levels. In the A.D(S) scenario at unsignalized intersections, pupil diameter increased by an average of 0.09 mm at the moment of sudden braking at the stop line (Seg. U-2) due to approaching vehicles compared to the moment before entry (Seg. U-1). This change in size was confirmed to have a moderate effect size of approximately 0.4, as measured by Cohen’s d. The pupil diameter showed a gradual decrease in subsequent moments, indicating that the intensity of the startle response was relatively high at that particular moment. In contrast, the A.D(C) scenario showed a significant decrease in pupil diameter during the subsequent segments compared to the moment before entry (Seg. U-1). This is interpreted as the driving style of A.D(C), which anticipates approaching vehicles and naturally slows down, giving participants a sense of stability. These patterns of pupil diameter changes are shown in Figure 12 and Figure 13.

Participants’ startle responses were more clearly observed in the pedestrian crossing situations. As shown in

Table 10. Results of user startle response ratings and physiological change analysis at crosswalk sections (n = 30).

Categories		A.D(S)				A.D(C)
		Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d	Mean Value	S.D.	Mean Difference (ρ-Value) *	Cohen’s d
Self-rating of startle response (5 scale)		3.87	-	-	-	1.44	-	-	-
EMG (arm) (0–1 n.u.)	Seg. C-1	0.366	0.205	-	-	0.395	0.161	-	-
	Seg. C-2	0.472	0.192	0.106 (0.007) **	0.536	0.417	0.173	0.022 (0.573)	0.106
Pupil diameter (mm)	Seg. C-1	3.827	0.637	-	-	3.936	0.667	-	-
	Seg. C-2	4.075	0.662	0.248 (0.001) **	0.961	3.815	0.654	−0.121 (0.001) **	0.717

* Numerical difference between the current segmentation and the first segmentation. ** ρ < 0.05.

, participants reported relatively high startle levels of 3.87 points in the A.D(S) scenario. This subjective rating was objectively verified by physiological measurements, which showed significant changes, with EMG (arm) measurements increasing by 0.106 and pupil diameter increasing by 0.248 compared to the moment before the event occurred (Seg. C-1). This change can be confirmed as a large effect size of approximately 0.96, which indicates a high level. In contrast, participants reported relatively lower levels of startle in the A.D(C) scenario, which was consistent with physiological measures showing a 0.121 decrease in pupil diameter compared to pre-event levels. These patterns of pupil diameter changes across segmentations are illustrated in Figure 14 and Figure 15.

4. Discussion

Automated-driving technology is rapidly becoming mainstream through initiatives like self-driving taxis and pilot public transportation programs. Beyond technical sophistication, securing user trust and acceptance has emerged as a critical challenge. Essential functions such as motion planning and decision-making for safe automated driving must be implemented in ways that users can trust and accept without anxiety across various traffic environments [51,70]. Research indicates that individuals who experience dangerous situations during automated driving show decreased trust, suggesting that high acceptance cannot be expected until safe automated driving is guaranteed [71]. Additionally, studies show that higher perceived risk by users can diminish trust in the technology regardless of system performance [72], highlighting the importance of eliminating unstable elements in automated-driving technology.

The main implications of this study are as follows. First, this study identified the relationship between the emotional and physiological responses of users in an unstable automated-driving situation. Human-centered automated-driving research has focused on evaluating the reliability of automated-driving behavior in various traffic situations [70,71,72]. In addition, studies have been conducted to evaluate how much anxiety or risk users feel during automated driving [28,29,30,49,50]. Previous studies evaluated users’ responses to dangerous behaviors such as sudden braking or sharp turns of automated-driving vehicles, and physiological changes, such as increased heart rate and dilated pupils, were observed in unrecognized dangerous situations.

This study is significant in that it analyzed the moments when frustration and startle response felt by users in automated-driving situations are maximized. Specifically, the frustration reported by participants was correlated with increased EMG (leg) activity, and the startle response in sudden braking situations was related to pupil dilation. The physiological responses of the participants showed greater changes in sudden vehicle-braking situations than in continuous frustration. This is consistent with previous research findings [29] showing a decrease in heart rate and dilated pupils in sudden situations, suggesting that the frustration and surprise reactions experienced in uncontrollable situations can cause physiological changes.

However, there are limitations in interpreting these physiological responses as simple cause-and-effect relationships. The highway merge-section case well illustrates this complexity. While participants reported high levels of frustration in this section, their physiological responses showed unexpected patterns. Specifically, EMG measurements showed a greater increase in arm activity rather than leg activity when driving in the acceleration lane, similar to startle responses at crosswalks. This suggests that the emotions experienced by participants in merge sections may have been more complex than simple frustration, emphasizing the need for in-depth research on users’ physical and psychological responses in various automated-driving situations.

Second, individual preferences should be considered when evaluating automated-driving behaviors in different traffic situations. Just as preferences for automated-driving vehicles differ by gender, age, education, and income [73,74], preferences for automated-driving styles also show significant individual differences [51].

Although this study failed to present significant statistical analysis results due to the small sample size, some differences according to gender and age group could be identified in the analyzed trends. In terms of frustration evaluation, the electromyography (EMG) changes were greater in women (difference between Seg. U-1 and U-3: 0.164) than in men (difference between Seg. U-1 and U-3: 0.065) at the uncontrolled intersection, and people in their 40s and 50s (difference between Seg. U-1 and U-3: 0.163) tended to have larger pupil diameter in situations of surprise than people in their 20s and 30s (difference between Seg. U-1 and U-3: 0.027). This may mean that women may be more nervous or have more sensitive physical reactions in automated-driving situations [49], and that people in their 40s and 50s are relatively less aware of dangerous situations than those in their 20s and 30s [50]. This suggests that personal factors such as gender and age may affect the evaluation of automated driving, and that research on personal factors is necessary for the development of automated-driving technology.

In addition, when we looked at the preferences of individual participants in detail, we found people who had different and heterogeneous evaluations from the majority. For example, only four participants rated frustration below 2 points in the roundabout section of the A.D(S) scenario, as shown in Figure 16. These participants tended to be more tolerant of waiting for gap acceptance or prefer safer driving behaviors. Conversely, in the A.D(C) scenario, five participants rated frustration above 4 points in the same section, indicating a preference for automated driving even with shorter gaps. As people show lower trust in automated driving when they experience dangerous situations [71], if the level of perceived risk varies among individuals, it will be difficult to ensure high trust and acceptance of automated-driving technology through universal automated-driving mechanism technology alone. Therefore, while prioritizing the implementation of commonly preferred driving behaviors when developing automated-driving operation mechanisms, it is also necessary to develop technology that allows for the selection of driving styles that reflect individual preferences.

These research findings suggest that adaptive systems capable of accommodating individual characteristics and preferences are necessary for the successful commercialization of automated-driving technology. Future development of motion-planning and decision-making algorithms should implement customizable driving modes that can be adjusted according to user preferences, rather than a uniform approach, to create an automated-driving environment in which people do not feel anxious. For example, just as adaptive cruise control provides systems optimized for individual driving propensities (conservative, moderate, and aggressive) [75], if automated driving also allows users to select driving styles optimized for individual propensities, people may feel less anxiety and have greater trust in automated-driving technology.

5. Conclusions

This study conducted experimental research using a virtual environment-based driving simulator to evaluate users’ negative physical and psychological responses (frustration and startle) in sections where unstable driving may occur during automated driving. Using simulation software, we implemented a virtual road environment with sequential driving on urban roads and highways, and reproduced situations known to be challenging for stable automated driving: roundabouts, highway merge sections, unsignalized intersections, and pedestrian crossings in blind spots. Based on this, we implemented both the A.D(S) scenario and the A.D(C) scenario, which enable stable driving, and comprehensively analyzed the participants’ negative responses through qualitative evaluation and physiological measurements (EMG and pupil diameter).

The analysis results showed that users experienced high levels of frustration due to long waiting times and strong startle response during sudden braking in automated-driving situations. These psychological reactions were objectively verified by physiological changes. Specifically, EMG measurements of leg muscles increased with longer waiting times, while pupil dilation and increased EMG measurements of arm muscles were observed in startle-response situations. This suggests that users may react negatively to automated-driving behaviors that differ from their expectations, potentially reducing trust and acceptance of automated driving.

However, as a preliminary study conducted with 30 participants, this research has limitations in generalizing negative psychological responses in unstable automated-driving situations. In particular, the relatively high proportion of men among the analyzed samples and the failure to obtain a balanced sample of people from different age groups mean that the possibility of biased results cannot be completely ruled out. In addition, the automated-driving scenarios implemented in this study simulated behavioral patterns based on publicly available data, which require further verification in experimental environments using actual automated-driving algorithms. Therefore, follow-up studies that address these limitations are needed to advance these research findings.

While the development of automated-driving technology and infrastructure is actively progressing worldwide, user-focused research is still at a relatively early stage, with significant potential for development. Ensuring user confidence in different traffic environments is essential for achieving practical results in automated-driving technology, which requires further research on user anxiety and acceptance in automated-driving situations.