Shared mobility is evolving as cities around the world repurpose traditionally “car-centric” public spaces to accommodate a variety of road users. Several challenges emerge for policymakers and researchers alike in developing a shared mobility ecosystem, including the need for multimodal integration and accessibility for all. With the advancement of automated vehicle (AV) technology, shared mobility services may provide important alternatives to conventional transportation and have the potential to alter the way in which people move in and around cities. While shared automated mobility has a role to play in the future of transportation, much research on its usage, impact on travel behavior, traffic congestion, and environmental impact requires exploration (Stocker and Shaheen, 2017). Another challenge that remains with respect to shared automated vehicles (SAVs) is developing a coherent understanding of how users will trust such systems (Mittendorf, 2017; Hartl et al., 2018) so that strategies for effective calibration of user trust can be developed. Moreover, most SAVs will be used on an “as-needed” basis, which may have another impact on the trust and adoption of the mobility. It is therefore incumbent upon researchers to study users' trust in different types of automated mobilities as well as how this trust may change when switching from one automated mobility type to another. In the next section, constructs of trust in automation are discussed.

Trust in automation has been studied for decades, and yet it remains a challenging topic for the wider research community. Persistent questions of how trust can be defined, established, and measured are areas of interest among researchers. One of the most accepted theories of how trust is established is given by the dyadic model of trust proposed by Mayer et al. (1995). The model defines trust as “the willingness to be vulnerable to the actions of others based on the expectation that the other will perform a particular action important to the trustor irrespective of the ability to monitor or control”. According to this model, trust depends on individual propensity (or general willingness) to trust others and the trustworthiness of the party to be trusted (trustee). A person's trust propensity results from different developmental experiences, personality types, and cultural backgrounds and determines how much a person trusts a trustee prior to their having any knowledge of that individual (Hoff and Bashir, 2015). The model establishes that trust is based on an individuals' ability to trust, benevolence to the extent to which a trustee is perceived to want to do good, and the integrity with which the trustee consistently adheres to a set of principles that the trustor finds acceptable. The ability to take risks is then the behavioral manifestation of the willingness to be vulnerable, i.e., the outcome of trust.

While this model establishes a strong working definition of the theoretical construct of trust, Lee and See (2004) helped further strengthen the understanding of how trust is established when humans interact with automation technology. They distilled the model of trust proposed by Mayer et al. (1995) into three dimensions: performance, process, and purpose. Performance seeks to define the current and previous operation of the automated system and how it comprises characteristics such as reliability, competency, and ability. Process describes how the automation matches the operators' objectives, which matches the integrity aspect. Finally, purpose describes the developers' intention to possess a positive orientation toward the operator, which aligns with the benevolence in Mayer et al. (1995). The comparison of these models has been described in great detail by Körber (2019) and forms the basis of our understanding of trust in automation.

Körber (2019) argues that trust is characterized by a static individual component, wherein humans rely on their inherent propensity to trust, depending on individual differences such as developmental experiences, personality type, and cultural backgrounds. This theoretical approach seeks to combine the constructs of interpersonal trust as well as trust in automation, which considers individual differences. The premise of this approach is to consider how reliable performance, predictable outcomes, and positive intentions of developers are factors of perceived trustworthiness, while prior experiences with automation (familiarity), individuals' propensity to trust in automation, and their general trust in automation act as moderating factors that influence perceived trust. Thus, trust in automation is defined as: “the attitude of a user to be willing to be vulnerable to the actions of an automated system based on the expectation that it will perform an action important to the user, irrespective of the ability to monitor or to intervene.” This definition implies that trust is more than the act of trusting performance, but rather a general attitude that the human is willing to be vulnerable based on their inherent state of trust (Mayer et al., 1995). In this paper, we adapt a survey developed by Körber (2019) and analyze each participants' trust along the dimensions italicized above. In the next section, individual preference in automated driving style is discussed.

While most autonomous vehicles (AVs) exhibit a default driving style, the default may not provide the most comfortable user experience for every individual. The mismatch between the AV's driving style and the user's expectation could result in lower trust in automation, thereby impeding the successful adoption of different SAVs (Trende et al., 2019). One approach toward fulfilling the expectation of the user is by tailoring the AV to match the style in which they would like to be driven. Indeed, several studies have demonstrated how a specific driving style can help improve the user's perception of an AV and contribute toward building trust (Ekman et al., 2019). It has been shown that fulfilling the drivers' maneuver preferences can lead to better comfort for them in an automated vehicle (Bellem et al., 2018). In fact, this level of personalization is not a new concept and has been approached by adopting preferences, patterns, styles, and skills of drivers in AVs (Hasenjager and Wersing, 2017; Natarajan et al., 2022).

Based on this evidence, we explicitly consider driving style preference in the proposed dual mobility experiment. Furthermore, we ensure that the driving style is consistent throughout nominal (standard) driving events and during “conflict events” in which the automated mobility demonstrates advanced decision-making while ensuring the safety and reliability of the vehicle. In the next section, we present the complete experiment design along with the study procedure and analysis methods used.

While promising, a critical challenge facing the adoption of a MaaS ecosystem is that many users may have infrequent or intermittent interactions with a variety of different shared vehicles and mobilities. It is well-accepted that user trust is critical for successful adoption of automated technologies (Lee and See, 2004; Parasuraman et al., 2008; Chen et al., 2010; Hancock et al., 2011; Cho et al., 2015). Therefore, it is necessary to understand users' trust in automation while experiencing new modes of automated mobility. More specifically, for a multimodal transportation ecosystem, it is critical to understand how trust in one automated mobility affects the trust in another automated mobility. Existing trust transfer theory in e-commerce suggests that trust can be transferred from one trusted source to another unknown target if there is a specific association between them (Doney and Cannon, 1997; Stewart, 2003). However, such a theory has not been studied in the context of interactions with automated vehicles. In particular, given the dynamic nature of trust (Cho et al., 2015) during a continuous interaction with automated vehicles, it is possible that the existing trust in one mobility not only influences the initial trust in another mobility but also influences how trust evolves in the latter mobility.

This raises the following research questions.

To address these questions, a novel dual mobility simulator experiment was designed that elicited participants' trust responses in two different automated driving (AD) systems. The two mobilities were a car and a sidewalk mobility. The car enabled with AD was a regular passenger car, whereas the sidewalk mobility was a personal-automated mobility typically used in urban environments which share space with pedestrians. Throughout the experiment, responses to trust questionnaires helped gauge how participants' trust changed with varying drives. Participants then shared their thoughts in a semi-structured interview at the conclusion of the experiment.

Our prior work showed (through a limited quantitative analysis of a subset of participants) whose trust transfers across mobility types based only on baseline and tutorial measurements (Mehrotra et al., 2023). In the previous paper, findings confirmed that aggressive style participants had a higher reliability when moving from car to sidewalk mobility. Additionally, it was found that the tutorial helped increase trust of aggressive participants in sidewalk mobility. In this paper, we expound upon the prior work by considering additional research questions regarding trust in automation, considering quantitative data from all experiment participants, and providing a rigorous thematic analysis of the qualitative data collected via semi-structured interviews with each participant.

The goal of this research is to evaluate how trust evolves in one automated mobility vs. another automated mobility with no prior interaction with either of the mobilities, and whether prior experience in an automated car influences trust in a novel automated mobility (that is not a car).

All participants completed two experimental sessions (one on each mobility) in a single day. For logistical reasons related to changing the mobility platform from car to sidewalk mobility and vice versa, a 3-h gap between the sessions was observed. The driving style of the mobility (defensive or aggressive) was chosen based on the participants' responses in the screening survey, as described in Section 2.5.1. This ensured that participants had the best chance to increase their trust in the mobilities.

In each experimental session, participants experienced three simulated drives: a tutorial drive, a standard drive, and a proactive drive. The three drives are explained in detail below, using terminology from Table 1.

A consistent approach to automation transparency was employed throughout the experiment to foster situational awareness and draw attention to proactive behavior. Two forms of transparency were used. First, augmented reality (AR) bounding boxes were used to highlight pedestrians, traffic signals, and other vehicles. When a traffic signal was highlighted, the color of the bounding box matched the color of the traffic signal. Second, at each conflict event, the system played a chime sound, followed by an audio (text-to-speech-based) notification. For example, for the “stale green light” conflict event shown in Figure 1, the speech notification for the aggressive proactive behavior said, “yellow light ahead; continuing” while that for the defensive proactive behavior said, “yellow light ahead; stopping.” To emphasize the color change of the signal, the AR bounding box highlighting the traffic light changed color from green to yellow.

Forty-eight participants were recruited for the study from and near San Jose, California. They were recruited through Craigslist jobs, social media, and a university research experience pool. Each applicant completed a screening survey in order to participate in the research. All participants were required to (1) be legally allowed to drive in the United States; (2) be 18 years of age or older; (3) have no self-reported hearing impairment; (4) have normal or corrected-to-normal vision using contact lenses (glasses could not be accommodated due to the virtual reality headset); and (5) not be easily susceptible to motion sickness. Additionally, a COVID-19 questionnaire was included to prevent anyone infected from participating. All 48 participants who completed the study were compensated $125; participants who failed to complete the study due to motion sickness or other constraints were compensated $25. The study was approved by the Institutional Review Board at San Jose State University.

A wide field-of-view (FoV) is necessary for a driver to observe events at an intersection; therefore, a VR headset with a broad FoV was chosen (Goedicke et al., 2018). Specifically, a StarVR headset with a 210-degree FoV was used in the study. A motion base (MB-200 6-degree of freedom motion base by Cosmate Co., Ltd.) was used to create a high-fidelity simulation in which participants could feel the typical forces experienced in a vehicle. The car and sidewalk mobility platforms were mounted upon the motion base, as seen in Figure 2. The automated driving was simulated by replaying a past researcher's drive via the “Wizard of Oz” technique (Wang et al., 2017). The display and audio were rendered using Unreal Engine 4.24 (Epic Games, 2019) with AirSim (Shah et al., 2018) that consisted of a custom city environment including traffic lights, other vehicles, pedestrians, stop signs, and roundabouts. Each drive was pre-recorded as a stereoscopic 360-degree video with a resolution of 4096 × 4096 along with the vehicle's position, orientation, and steering input. The recorded drives were played on the StarVR headset using a custom software developed in Unity (Unity Technologies, 2019) that synchronously controlled the motion base and the steering wheel to mimic the vehicle's behavior in the recording.

For this experiment, there were 3 between subject factors: (1) Driving style (Aggressive or Defensive) (2) Order of mobility (Car to sidewalk or Sidewalk to Car), and (3) Order of treatment in the afternoon session (Tutorial → Standard → Proactive or Tutorial → Proactive → Proactive). A 2 × 2 × 2 between-subject experiment was considered for the experiment design. Twenty-four participants experienced the aggressive driving style, whereas the other 24 participants experienced the defensive driving style. For each driving style, 12 participants experienced the car (morning session) first, followed by the sidewalk mobility (afternoon session). The remaining 12 participants experienced the sidewalk (morning session) first, followed by the car mobility (afternoon session). In the afternoon sessions, for all driving styles and mobilities, 6 participants experienced the “standard” drive followed by the “proactive” drive. The other 6 participants experienced “proactive” drives twice. The complete study design is summarized in Figure 3. We considered ways to control for several confounding variables for the experiment: (1) lack of familiarity with the simulator environment was mitigated by providing tutorial drives to get participants accustomed to the environment; (2) to ensure attention, the participants were encouraged to speak throughout the experiment describing how they felt, which kept them engaged with the study; and (3) to ensure participants were familiar with traffic conventions, only participants with valid driver's license was included.

The study procedure consisted of four parts: a screening survey, a morning session with one automated mobility followed by a gap, an afternoon session with a second automated mobility, and finally a semi-structured interview. Additional details are provided in the following subsections.

Thirty-three of forty-eight participants reported whether or not they initially trusted their mobility during session 1 (see Table 5). Six participants, for both the car and sidewalk mobility, reported an initial state of trust in their respective mobility. However, ten participants reported an initial state of distrust in the sidewalk mobility compared to seven car mobility participants. This section focuses on answering the research question of how trust evolves in each of the mobilities beyond these initial perceptions during session 1.

Thirteen of twenty-four participants who experienced the car mobility during their first session said that their trust transferred to the sidewalk mobility during their second session. Four of those twenty-four participants said that their trust did not transfer to the sidewalk mobility from the car. Seven participants either did not comment on whether their trust transferred or not or their semi-structured interview audio was not audible.

Of the thirteen participants whose trust did transfer from the car mobility to the sidewalk mobility, several shared their experience and why they felt that their trust transferred. As Participant 6 explained, “I think after going in the car, I got to trust it and I got what I was doing, so I feel like with a scooter I expected kind of the same thing, like how it was gonna stop and speed up and how it was gonna react to people and other cars.” Two participants reported they had positive expectations of the second mobility because they assumed both mobilities were created by the same developers. Participant 30 stated, “If I did the car, and it seems like it was trustworthy, the same people are developing this one. So it must have the same principles.” Some participants suggested that they had a general idea of how the second mobility would behave from a technological perspective. Participant 1 had a very good experience with the car mobility, so they said referring to the sidewalk mobility, “if it's by the same developers and kind of works the same way, then I really have a higher level of trust and confidence in it.”

On the contrary, four participants had a different view that trust did not transfer across mobilities. Participant 11 responded, “I think it was pretty independent” when asked if trust in the car mobility from session 1 influenced their trust in the sidewalk mobility in session 2. However, Participant 11 did confirm that overall they felt safer in the car. Participant 36 reiterated the sentiment shared by Participant 11, “I don't think it (the car mobility) influenced at all like how I viewed [the sidewalk mobility]; it felt like two completely different tasks.” Participant 34 saw the two experiences as separate due to the physical size of the mobilities and their potential of hurting someone. Lastly, Participant 32 felt like the experience in session 1 with the car mobility is in a more structured environment whereas on the sidewalk it is “too spontaneous” due to pedestrian variability; therefore, it felt like the two experiences were not interdependent.

To analyze the effect of the interaction with an automated car on trust in the automated sidewalk mobility, we compare the sidewalk mobility responses for the participants who interacted with the sidewalk mobility in session 2 to that of the participants who interacted with the sidewalk mobility in session 1. This allows us to evaluate whether (or not) the prior experience with an automated car affected the evolution of trust in the sidewalk mobility. To mitigate the effect of drive type, we only consider the survey results of the 6 participants who had the same standard-proactive drives on the sidewalk mobility during session 2 for comparison to that of the 12 participants with the sidewalk mobility in session 1. To investigate the statistical significance between the responses of the two groups of participants for a given drive type, we perform a two-sample t-test for each of the drive types for all dimensions of trust. For example, we compare BaseSMSes1 vs. BaseSMSes2, TutSMSes1 vs. TutSMSes2, StdSMSes1 vs. StdSMSes2, and ProSMSes1 vs. ProSMSes2 for the dimension intention of developers.

Figures 8, 9 show the changes in the dimensions of trust across the drives for defensive and aggressive participants, respectively. For defensive drivers (Figure 8), the two-sample t-test results were significant for intention of developers. These results are shown in Table 6. This result demonstrates that prior experience with an automated car leads to a significant trust increase with respect to the intention of the developers who made the automation, but no significant difference in trust in automation of the actual system. Figure 9 illustrates results for participants who experienced aggressive AD behaviors. None of their two-sample t-test results were significant, implying that aggressive driving style of cars did not influence the participants' trust in the sidewalk mobility.

During the semi-structured interview, many participants discussed trust in different automated mobilities. Apart from trust, one emergent theme from the interviews was the importance of transparency. For this study, bounding boxes highlighted other vehicles and pedestrians on the same side of the road. Additionally, vehicles provided voice alerts during proactive events. 14/48 participants expressed positive opinions on voice alerts, as it helped them understand the mobility's current and future actions. One participant stated, “[the voice assistance] is double assurance that the system is actually. . .keeping up with everything; and it's not going to run over someone.” However, some participants expressed concerns about the promptness of the voice assistance, and reported that it came too late in order to be helpful. 10/48 participants found the voice alerts barely perceptible, easily ignored, or redundant. A participant stated “Maybe at the beginning [it was helpful], like it helped me know what was going to happen. But at the end. . .I knew what was going to happen, so it was very unnecessary.” These observations highlight the importance of automation using variable or adaptive transparency. When users experience automation initially, greater amounts of transparency are necessary to help increase their trust; however, as expressed by some participants, a continuously high level of transparency can become redundant or distracting.

In summary, the first research question explored how human trust evolved in one automated mobility vs. another automated mobility with no prior interaction with either of the mobilities. Findings suggest that while trust increases over time, the effect of mobility was not significant. The second research question explored how prior experience in an automated car influence trust in a novel automated mobility (that is not a car). Findings suggest that only the intention of developers was found to be significant.

We found that during an initial session with either a car or sidewalk mobility, all five dimensions of trust increased similarly. These dimensions are reliability, predictability, intention of developers, propensity to trust, and trust in automation. This suggests that there may not be differences in how trust builds in an automated car vs. an automated sidewalk mobility. Additionally, we observed that participants who preferred a defensive automated driving style had a significantly higher amount of trust in the intention of developers while using the sidewalk mobility during their second session vs. their defensive counterparts using the sidewalk mobility during their first session. Overall, participants confirmed in their semi-structured interviews that their trust increased from the first mobility to the second.

Another important finding from this work is that vehicle type had no effect on any dimensions of trust as shown in Section 3.1.3. This result suggests that users' trust can grow and evolve across different mobilities even when they are unfamiliar with a particular mobility. This has important implications for several stakeholders in a future MaaS ecosystem, in which it is expected that users will frequently interact or transition between different mobility types. As mentioned in Section 3, some participants reported that their trust transferred from the first mobility to the second due to an assumption that the mobilities were made by the same “company”, or “developers”, or use the same “software”, “system”, or “artificial intelligence”. This suggests that brand loyalty to automakers, which certainly influences consumers as they make personal vehicle purchasing decisions (Jørgensen et al., 2016; Nadzri et al., 2016), may also extend to users' trust in automated mobilities, even when the “vehicle” is not a conventional car or other types of automobile.

Interestingly, we also found that certain drive types had a significant effect on all dimensions of trust. Specifically, there was a statistically significant effect on all dimensions of trust for participants when we compared their baseline reported trust to that after they experienced the standard drive. Drives with defensive actions had the most consistent and significant effects on trust. Drives with aggressive actions decreased trust in the intention of the developers, likely due to what could be perceived as dangerous actions at certain events. These results suggest that trust can increase or decrease in accordance to the type of experience a user has in a given drive. Similar to purchasing a new car, users may need opportunities to “test drive” and positively experience new forms of mobility. Education campaigns have already been developed for e-bikes and e-scooters in cities including New York City in response to increasing safety concerns (Hu and Marcius, 2021). Automated mobilities could potentially exacerbate safety concerns when sharing sidewalks and roadways with electric or manually-powered personal mobility devices. Therefore, considering a future MaaS ecosystem, cities in particular may need to consider training or education programs for residents on varied mobility types, both to improve safety and adoption.

The trends shown in Figure 6 (aggressive) indicate that the baseline levels of trust in the sidewalk mobility are similar to the baseline in Figure 5 (Defensive). What differs are the trends seen in the car mobility. With the car being the more familiar form of mobility, the human's trust dynamics when experiencing aggressive actions exhibit similar trends to an unfamiliar form of mobility—the sidewalk mobility. This finding indicates that the aggressive driving experience in a car can be likened to that of something unfamiliar and thereby influence trust. Interestingly, although not significant, we see trends from Figure 6 that imply individuals who preferred aggressive driving showed an increased willingness to rely on and trust the car despite the fact that their trust in the car's predictability decreased when they encountered sequential proactive behavior. This suggests that proactive automated driving can increase trust and reliance in a car mobility which implies “faith,” or a higher level of trust, in the automated car. The introduction of “conflict” events in proactive drives was a novel aspect of the experiment as it served to stimulate changes in trust in ways that were more consistent with real driving scenarios. Therefore, this finding is especially encouraging in that it shows the participants may trust more advanced decision-making on the part of an automated vehicle.

Lastly, the use of voice prompts to increase automation transparency was helpful and reassuring to some participants. However, participants reported that while the transparency was appreciated initially, it became a distraction as participants became more familiar with each mobility. This provides a strong motivation for the use of adaptive transparency (Akash et al., 2020b) that is sensitive both to individual user's preferences as well as time.

It is likewise important to highlight some limitations of the work presented here. First, certain aspects of the vehicle simulators could be changed to improve ecological validity. To mitigate the negative effects of simulator glitches, each drive was pre-recorded. While this ensured a consistent experience for each participant, it did not allow participants to actually takeover control of the vehicle. Instead, we were limited to only measuring “intent to takeover”. The second aspect that should be revisited is the smoothness of each mobility. The car simulator intentionally had a smoother ride than the sidewalk mobility for realism. However, some participants reported that the dynamics of the sidewalk mobility were too bumpy during some portions of the drives. This could have introduced confounds in the users' experience. Therefore, the settings on each simulator should be adjusted in future experiments. For observing individual differences and demographic factors, future studies should consider controlling for differences in age, gender, and culture to understand if those factors influence trust across mobilities. While countermeasures in our experimental design did not introduce bias due to participant fatigue and affect, future studies could account for fatigue and affect for measuring trust in automation. Finally, a limitation of this study is the small sample size of participants who experienced the sidewalk mobility in session 2 that we could compare with those who experienced it in session 1. To validate the differences observed in this comparison, follow-up experiments with a larger sample size should be conducted.

Nevertheless, despite these limitations, this dual mobility experiment and associated analysis represents the first investigation of trust transfer across different automated mobility types and serves as a foundation for future work aimed at enabling a shared mobility future.