Why Do Users Switch from Ride-Hailing to Robotaxi? Exploring Sustainable Mobility Decisions Through a Push–Pull–Mooring Perspective

Abstract

Robotaxi services represent a major step in the commercialization of autonomous driving, offering efficiency, consistency, and safety benefits. However, despite technological advances, their large-scale adoption is far from guaranteed. Most urban users already rely on mature ride-hailing platforms such as Didi and Uber, making the real behavioral question not whether to adopt Robotaxi, but whether to migrate from existing services. Prior studies based on TAM, UTAUT, or trust models have primarily examined users’ initial adoption decisions, overlooking the substitution behavior that better captures how people shift between competing mobility services in real contexts. This study addresses this gap by applying the Push–Pull–Mooring (PPM) framework to examine users’ migration from ride-hailing to Robotaxi services, based on survey data collected from 1206 respondents across four Chinese cities (Beijing, Shanghai, Guangzhou, and Wuhan). The model was tested using structural equation modeling and multi-group analysis (SEM–MGA). Push factors reflect negative experiences with ride-hailing, including social anxiety and insecurity caused by drivers’ behaviors; pull factors emphasize Robotaxi’ autonomy and service reliability; while mooring factors capture habitual ride-hailing use and perceived Robotaxi risk. Findings indicate that push and pull factors significantly promote migration intentions, whereas mooring factors hinder them. Among all factors, perceived risk exerted the strongest negative effect (β = −0.36), underscoring its critical role as a barrier to Robotaxi migration. Gender differences are also evident, with women more sensitive to risks and men more influenced by reliability. By situating adoption within a migration context, this study enriches high-risk innovation theory and offers practical guidance for designing gender-sensitive and user-specific promotion strategies.

1. Introduction

With the rapid iteration of autonomous driving technologies, autonomous taxis (Robotaxi) have become a focal point of attention in both academia and industry [1]. Commercial pilot projects have already been launched in various regions worldwide, including cities in China, the United States, and Europe [1,2]. According to McKinsey’s forecast, the Robotaxi market is expected to reach tens of billions of dollars by 2030. Its potential impact extends beyond transforming transportation modes to reshaping retail and service markets [3,4]. In practice, the number of Robotaxi vehicles and pilot cities has increased rapidly. For example, by 2024, over 20 Chinese cities—including Beijing, Wuhan, and Shanghai—had authorized public Robotaxi operations, while Waymo and Cruise have expanded to more than 10 cities in the United States [3,5]. For this emerging industry, user acceptance and switching behavior are pivotal to achieving large-scale commercialization. Consequently, understanding how consumers make behavioral choices toward Robotaxi has become an issue of pressing concern to both scholars and practitioners.

Existing studies on Robotaxi adoption have primarily drawn on technology acceptance and trust-based models [6,7,8], identifying several technical and psychological drivers of user acceptance [9,10,11]. However, these frameworks conceptualize adoption in isolation and fail to account for the competitive context in which users evaluate Robotaxi relative to existing mobility options. While these insights are valuable, they are predominantly grounded in a “vacuum adoption” assumption—as if users decide to embrace a new technology from a zero starting point. In other words, user acceptance is conceptualized as the adoption of a novel service in isolation, rather than as a decision made against existing alternatives. In reality, when Robotaxis enter the marketplace, users are already accustomed to and frequently rely on another mature mobility option—ride-hailing. Due to its transparent pricing, flexible service, and everyday convenience, ride-hailing has become the dominant travel choice in many urban settings [12]. This means that Robotaxis are introduced within a competitive market, where they must directly compete with established ride-hailing services. Hence, the real question for users is not whether to accept a new technology, but whether to switch from ride-hailing to Robotaxi. This migration logic has been largely neglected in the existing literature, leaving unanswered why some users would abandon an established alternative while others would not, even when trust in Robotaxi is enhanced.

Beyond this central gap, two additional areas merit deeper investigation. While many studies have examined safety and risk, they often treat them as broad, undifferentiated constructs [13]. When distinctions are made, they are seldom integrated into a migration context or operationalized in a way that explains switching behavior. On the one hand, there are human-related risks associated with driver behavior and social interactions, such as traffic violations, harassment, or communication stress [14]. On the other hand, there are technology-related risks arising from driverless systems, such as technical failures or data breaches [15,16].Previous research on human–automation interaction has also shown that users’ perceived risks in autonomous mobility mainly stem from social–behavioral (human-related) and technological sources [10,17]. Building on this foundation, this study explicitly distinguishes these two sources to examine how they differentially shape migration from ride-hailing to Robotaxi. Failing to distinguish these sources makes it difficult to explain why some users prefer to tolerate driver-related risks rather than face technological uncertainties. Second, most existing studies adopt an average-effects lens, overlooking heterogeneity across user groups, particularly gender differences [18]. Furthermore, gender differences in risk sensitivity and technology evaluation have been consistently observed in mobility research [19,20], suggesting that user heterogeneity plays a critical role in shaping migration motivations but remains underexplored in the Robotaxi context [21,22]. Collectively, these limitations reveal that the existing literature lacks an integrated understanding of how migration logic, differentiated risk perceptions, and user heterogeneity jointly influence users’ switching behavior from ride-hailing to Robotaxi.

Building upon this recognition, the present study therefore addresses the following core research question: Why do users choose (or refuse) to migrate from ride-hailing to Robotaxi when both options are available? To answer this question, we adopt the Push–Pull–Mooring (PPM) framework and situate Robotaxi adoption within a genuine migration context. Unlike TAM or UTAUT, which emphasize single-point adoption logic and are often supplemented by trust and risk perspectives [23,24], the PPM framework captures three sets of forces simultaneously: push forces that drive users away from existing services, pull forces that attract them toward new services, and mooring forces that hinder switching [25,26]. This multidimensional perspective not only aligns with the competitive coexistence of Robotaxi and ride-hailing but also provides a more explanatory lens for understanding users’ psychological trade-offs across different risk sources and demographic groups. Accordingly, the PPM framework offers a theoretically stronger and more practically relevant entry point for examining Robotaxi adoption.

Based on this perspective, our study makes three primary contributions. First, we move beyond the isolated adoption perspective and the vacuum-based adoption assumption and, for the first time, conceptualize Robotaxi adoption as a migration decision in direct competition with ride-hailing, thereby enriching theoretical understanding of emerging service adoption. Second, we distinguish and operationalize human-related versus technology-related risks, demonstrating their roles in migration decisions, and in doing so extend the PPM framework within consumer behavior research, offering a new explanatory dimension for risk and user-experience studies. Third, through gender-based group analysis, we uncover heterogeneity in migration motivations, showing that men and women differ significantly in their risk perceptions and pull evaluations. These findings provide empirical evidence to inform the differentiated design and targeted promotion of Robotaxi services.

2. Literature Review

2.1. From Ride-Hailing to Robotaxi: User Research

Over the past decade, ride-hailing has rapidly emerged as a dominant form of shared mobility worldwide [19]. Companies such as Uber and Lyft pioneered app-based ride-hailing in North America, while Didi in China, Grab in Southeast Asia, and Ola in India achieved large-scale expansion within a short period [27]. By offering transparent pricing, real-time tracking, and seamless payment, these platforms substantially alleviated the information asymmetries and service inconsistencies long associated with traditional taxi markets. As a result, ride-hailing has become part of the everyday routines of urban residents [28].

Academic research on ride-hailing has focused on two major streams. The first concerns user acceptance and usage motivations. Studies indicate that price advantages, service flexibility, and operational convenience are key drivers of adoption [29], while driver-rating systems and real-time feedback mechanisms enhance trust and overall satisfaction [12]. The second stream centers on risks and social implications. Although ride-hailing offers convenience, it also raises safety, privacy, and labor relations concerns [30]. For example, driver fatigue, harassment incidents, and communication barriers are frequently cited as factors undermining users’ sense of security and continued use [14]. Notably, female users exhibit heightened safety sensitivity, a finding consistently validated across studies. Overall, the literature highlights the value of ride-hailing within shared mobility but primarily addresses adoption and usage in isolation, with little attention paid to users’ comparative evaluations and migration toward new mobility services.

In parallel with the widespread adoption of ride-hailing, Robotaxis have recently entered the commercialization stage. Waymo and Cruise pioneered autonomous taxi services in the United States [31], while Baidu Apollo, AutoX, Pony.ai, and Didi launched demonstration operations in Chinese cities such as Beijing, Shanghai, and Shenzhen [10,32]. Pilot programs have also accelerated in parts of Europe and the Middle East, marking the transition of Robotaxi from “proof-of-concept” to “limited deployment” [33]. Unlike ride-hailing, Robotaxi serve not only as a transportation service but also as a critical consumer-facing application of autonomous driving technology.

Scholarly research on Robotaxi has developed along two principal lines. The first is the logic of technology acceptance. Numerous studies, grounded in TAM or UTAUT, examine how factors such as performance expectancy, effort expectancy, social influence, and facilitating conditions shape adoption intentions [6]. For example, users’ expectations of improved efficiency and time savings have been shown to positively predict adoption [34]. The second is the logic of trust and risk. Given their high degree of automation, Robotaxis raise trust and safety concerns, with system transparency, technical reliability, privacy protection, and risk perception identified as critical antecedents of trust and adoption [32,35]. Some studies also highlight Robotaxi’s distinctive experiential dimensions—such as autonomy, consistency, and technological novelty—that can enhance user excitement and satisfaction [36]. However, these works typically assume a single-point adoption process, implicitly treating Robotaxi adoption as a decision made from scratch, while overlooking users’ established habits and reference points shaped by ride-hailing.

In summary, existing research on ride-hailing and Robotaxi has generated important insights [32,37], but explanations of migration behavior between the two remain underdeveloped. First, mainstream studies perpetuate the “vacuum assumption,” framing Robotaxi as a standalone alternative without capturing the real-world decision of abandoning ride-hailing in favor of Robotaxi. Second, risk factors are often treated in an overly broad manner, with little differentiation between human-related risks (e.g., driver behavior, social interaction pressure) and technology-related risks (e.g., system failures, data breaches), thereby limiting understanding of users’ preference trade-offs between these risk types. Third, most studies adopt an average-effects perspective, neglecting heterogeneity across user groups—particularly gender differences. For instance, women tend to prioritize safety and privacy, whereas men emphasize efficiency and autonomy. Ignoring such divergence risks distorts interpretations of users’ migration motivations.

Therefore, while current research offers valuable insights into shared mobility services, its limitations prevent us from addressing a critical question: Why do users choose (or refuse) to migrate to Robotaxi when ride-hailing remains available? To address this gap, the present study situates Robotaxi adoption within a migration context and employs the Push–Pull–Mooring framework as an integrative lens to systematically analyze the drivers, attractors, and barriers influencing user migration decisions.

2.2. Push–Pull–Mooring (PPM)

The Push–Pull–Mooring (PPM) framework was first introduced by Lee to explain population migration [38] and has been widely extended to technology adoption and green mobility contexts in recent years [26,39]. It posits that migration is rarely the result of a single factor but is shaped by three categories of forces: push factors represent the negative pressures that drive individuals away from an existing option, pull factors denote the positive attractions of a new alternative, and mooring factors capture the constraints and barriers that hinder switching. Because PPM simultaneously accounts for driving, attracting, and restraining forces, it has been increasingly applied in consumer behavior and service research, where it is recognized as a more powerful explanatory tool for migration than single-point adoption models [25,40].

In information systems and digital services, PPM has been widely employed to explain user switching across platforms and services. For example, in studies of e-commerce migration, dissatisfaction and loss of trust in the original platform often serve as push factors, while superior functionality and personalized services of the new platform act as pull factors; shopping habits, sunk costs, and platform lock-in function as mooring forces [41,42]. Similar dynamics appear in research on social media and mobile applications, where users’ network ties and social capital strengthen mooring effects, indicating that migration is not only determined by the appeal of a new platform but also constrained by the stickiness of existing social environments [43,44,45]. Collectively, these studies show that PPM offers a systematic lens for capturing users’ psychological trade-offs under multiple choices and is well-suited for explaining cross-platform and cross-service switching [46].

More recently, the PPM framework has been introduced into the automotive domain, particularly to explain the transition from conventional fuel vehicles to new energy vehicles (NEVs). Research consistently finds that environmental concerns, policy incentives, and dissatisfaction with the costs and environmental burdens of fuel vehicles are typical push factors [47,48]. The economic, convenience, and technological advantages of NEVs represent pull forces [26], while entrenched driving habits, risk perceptions of new technologies, and high switching costs function as mooring forces [49]. These findings confirm PPM’s applicability to transportation mode choices and highlight that user migration is influenced by a combination of environmental, policy, experiential, and habitual factors. However, most of these studies are confined to “private vehicle energy substitution,” focusing on the shift from fuel cars to EVs, while systematic investigation of migration in shared mobility—particularly from ride-hailing to Robotaxi—remains scarce.

This research gap is theoretically and practically significant. Compared with private vehicle substitution, migration decisions in shared mobility are far more complex [50]. They involve not only technology acceptance and risk perception but also the reconfiguration of travel experiences, social interactions, and perceived safety. Previous Robotaxi studies, mostly grounded in TAM, UTAUT, or trust–risk models, have provided valuable insights into users’ acceptance and safety perceptions [6,51]. However, these models conceptualize Robotaxi adoption as an isolated decision rather than a comparative evaluation against existing mobility options. For example, human-related risks and social pressures in ride-hailing may act as push factors, whereas the autonomy, consistency, and technological novelty of Robotaxi may serve as pull factors [14,52]. Meanwhile, dependence on established travel habits, concerns over technical failures, and privacy risks may operate as mooring forces. Importantly, existing studies often overlook structural differences across user groups, particularly gender. Prior research shows that women prioritize safety and privacy in travel decisions, whereas men are more motivated by efficiency, autonomy, and novelty. Such distinctions are especially relevant in the Robotaxi context, where the removal of the human driver alters traditional sources of risk and comfort. Recognizing these gender-based differences is therefore essential for explaining why users vary in their willingness to migrate from ride-hailing to Robotaxi [9,53].Unlike these models, the PPM framework offers a more comprehensive and dynamic perspective by capturing not only users’ acceptance determinants but also the psychological mechanisms of behavioral migration. It enables the simultaneous consideration of dissatisfaction with existing ride-hailing (push), attraction to autonomous advantages (pull), and habitual or risk-based constraints (mooring), thereby providing a more realistic explanation of how users evaluate and switch between competing mobility options.

Building on the above literature, this study further clarifies how the key theoretical constructs correspond to the three dimensions of the PPM framework (Figure 1). Specifically, push factors are operationalized as users’ negative psychological responses toward human-driven ride-hailing, captured by social anxiety and human-induced insecurity, which reflect interpersonal unease and perceived safety concerns caused by driver behavior and social interaction. Pull factors represent the perceived attractions of Robotaxi, including perceived autonomy and perceived service reliability, which highlight the technological advantages, consistency, and sense of control offered by automation. Mooring factors refer to the psychological and behavioral barriers that resist switching, represented by habit—the inertia of prior ride-hailing use—and perceived Robotaxi risk, which encompasses uncertainty and potential loss in adopting an autonomous service. Together, these constructs concretely embody the three dimensions of the PPM framework and allow a systematic examination of users’ migration mechanisms from ride-hailing to Robotaxi. These six constructs were derived directly from prior findings on users’ emotional (social anxiety, insecurity), functional (autonomy, reliability), and behavioral (habit, perceived risk) responses in mobility transitions, representing the most salient mechanisms shaping migration from human-driven to autonomous services.

Therefore, building on and extending prior Robotaxi research, this study introduces the Push–Pull–Mooring (PPM) framework as a more comprehensive and integrative lens. Unlike earlier models that examine adoption as a single-point decision, PPM simultaneously captures the driving, attracting, and restraining forces underlying real-world behavioral migration. This allows the integration of previously fragmented findings on trust, risk, and user experience into a unified structure, offering stronger explanatory power for understanding how users switch from ride-hailing to Robotaxi.

In summary, although the PPM framework has been validated in information systems, service switching, and private vehicle energy transitions, it has yet to be systematically applied to explain the migration from ride-hailing to Robotaxi. This study introduces PPM into this emerging context to uncover the mechanisms through which push, pull, and mooring forces shape Robotaxi migration and to further explore how gender differences contribute to divergent migration motivations.

3. Model Conceptualization and Hypotheses Development

Drawing on the preceding literature review, this study employs the Push–Pull–Mooring (PPM) framework to explain the behavioral mechanisms underlying users’ migration from ride-hailing to Robotaxi. Within this framework, push factors capture users’ dissatisfaction and perceived risks associated with existing services, pull factors represent the attractiveness of the new service, and mooring factors denote path dependencies and perceived constraints that hinder migration.

This study extends prior applications of the PPM framework in three important ways. First, it differentiates between human-related risks and technology-related risks, thereby better reflecting the realities of mobility scenarios. Second, it incorporates autonomy and service reliability into the pull dimension to highlight the distinctive value propositions of Robotaxi. Third, it introduces the moderating effects of usage habits and perceived risk to account for heterogeneity in user migration decisions. Based on these extensions, the research model is depicted in Figure 1.

3.1. Push Hypotheses

3.1.1. Social Anxiety (SAN) and Switching Intention (SWI)

Social anxiety refers to the psychological discomfort or tension experienced by individuals during interpersonal interactions [32]. In the ride-hailing context, passengers are often involuntarily drawn into communication with drivers, such as confirming routes, engaging in small talk, or clarifying fares. These situations may generate unnecessary social pressure [14]. Prior studies indicate that social anxiety significantly influences users’ travel choices, with anxious individuals tending to avoid socially intensive environments and preferring travel modes that minimize interpersonal interaction [54].

By eliminating direct driver–passenger interactions, Robotaxis reduce such social pressures and enhance passengers’ sense of autonomy and privacy [55]. Accordingly, higher levels of social anxiety are expected to increase the likelihood that users abandon ride-hailing in favor of Robotaxi. Thus

Hypothesis H1a.

In the ride-hailing context, social anxiety (SAN) exerts a positive effect on switching intention (SWI).

3.1.2. Human-Induced Insecurity (HII) and Switching Intention (SWI)

Human-induced insecurity refers to the sense of unsafety or uncertainty that arises from drivers’ behaviors or personal characteristics during a trip. In ride-hailing, such insecurity is commonly triggered by risky or unlawful driving practices, including speeding, red-light violations, distracted driving (e.g., smartphone use), or emotionally charged interactions that make passengers feel threatened. Previous research has shown that unpredictability stemming from human factors constitutes a major source of negative passenger experiences [52].

The key advantage of Robotaxi lies in delegating driving tasks to algorithms and automated systems, thereby eliminating the variability associated with human drivers. This mechanism mitigates risks induced by human error and reduces the likelihood of unsafe driving, ultimately enhancing passengers’ safety expectations. Thus, it can be inferred that:

Hypothesis H1b.

In the ride-hailing context, human-induced insecurity (HII) positively influences switching intention (SWI).

3.2. Pull Hypotheses

3.2.1. Perceived Autonomy (PAU) and Switching Intention (SWI)

Perceived autonomy refers to the sense of control and self-determination that users experience when engaging with a service [56]. In the context of mobility services, autonomy manifests in passengers’ ability to preserve privacy, personal space, and decision-making authority during travel. Prior studies indicate that higher levels of perceived autonomy enhance users’ acceptance of emerging mobility modes [57].

In the Robotaxi context, the absence of a driver eliminates social interruptions, enabling passengers to enjoy greater freedom and privacy. Unlike ride-hailing, where the driver’s presence often generates a sense of being observed, thereby constraining behaviors such as resting, making phone calls, or engaging in leisure activities, Robotaxis provide a travel space free from social pressure. This environment not only satisfies individuals’ autonomy needs, as emphasized in self-determination theory [58], but also enhances psychological comfort and perceived control by reducing social presence [59]. Accordingly, it is hypothesized that:

Hypothesis H2a.

The perceived autonomy (PAU) offered by Robotaxi positively influences switching intention (SWI).

3.2.2. Perceived Service Reliability (PSR) and Switching Intention (SWI)

Perceived service reliability refers to users’ overall perception of the stability and consistency of a service [60]. In ride-hailing, variations in drivers’ driving styles, attitudes, and route selections can result in highly inconsistent experiences across trips, sometimes producing substantial uncertainty and unpredictability [14]. Prior research has shown that service reliability is a critical determinant of continued usage in mobility services [61].

By contrast, Robotaxi operations are governed by algorithms and standardized protocols, ensuring high consistency in route selection, driving logic, and trip execution [10]. For instance, Robotaxis strictly comply with traffic regulations, apply uniform acceleration and braking patterns, and deliver reproducible travel experiences. Such consistency not only alleviates concerns over service variability but also strengthens users’ sense of predictability and trust in the outcomes of their journeys. Hence, it is hypothesized that:

Hypothesis H2b.

The perceived service reliability (PSR) offered by Robotaxi positively influences switching intention (SWI).

3.3. Mooring Hypotheses

3.3.1. Habit (HAB) and Switching Intention (SWI)

Habit refers to an individual’s automatic behavioral tendency formed through repeated actions over time [24]. In the context of mobility services, frequent use of ride-hailing fosters habitual usage scripts, thereby reducing users’ willingness to explore alternative options. Prior research suggests that entrenched habits often serve as major barriers to the adoption of new technologies [62].

Beyond operational familiarity, ride-hailing habits also encompass embedded practices such as payment method binding, route preferences, reward points, and coupon systems, all of which gradually integrate ride-hailing into users’ daily mobility ecosystems. Such path dependence and cognitive lock-in make users more inclined to maintain the status quo rather than invest effort in adapting to a new mobility mode. Consequently, even if Robotaxis offer superior advantages, strong ride-hailing habits may significantly impede migration.

Hypothesis H3a.

In the ride-hailing context, user habit (HAB) negatively influences switching intention (SWI) toward Robotaxi.

3.3.2. Perceived Robotaxi Risk (PRR) and Switching Intention (SWI)

In the Robotaxi context, perceived risk primarily manifests as concerns about technological failures, system uncontrollability, and potential privacy breaches [32]. Prior studies demonstrate that perceived risk undermines user trust in new technologies, thereby substantially reducing adoption and migration intentions [63,64]. Accordingly, it can be inferred that the higher the perceived risk of Robotaxi, the lower the likelihood of switching.

Hypothesis H3b.

Perceived Robotaxi risk (PRR) negatively influences switching intention (SWI).

3.4. Moderating Hypotheses

3.4.1. The Moderating Role of Habit (HAB)

Habit not only exerts a direct effect on switching intention but may also moderate the relationship between push factors and switching intention. Prior studies suggest that habitual usage diminishes the extent to which dissatisfaction drives user behavior [65]. From the perspectives of status quo bias [66] and inertia theory [67], once users develop stable behavioral routines, they may continue using existing services despite experiencing social anxiety or human-induced insecurity. Accordingly, strong ride-hailing habits are likely to weaken the positive influence of push factors on migration.

Hypothesis H4a.

Habit (HAB) weakens the positive effect of social anxiety (SAN) on switching intention (SWI).

Hypothesis H4b.

Habit (HAB) weakens the positive effect of human-induced insecurity (HII) on switching intention (SWI).

3.4.2. The Moderating Role of Perceived Robotaxi Risk (PRR)

Perceived risk not only directly reduces switching intention but may also attenuate the positive effects of pull factors. According to the risk–return trade-off model [68], when users perceive high risks, they tend to adopt defensive cognition and become less responsive to service advantages. In the context of Robotaxi, even if autonomy and service reliability are provided, users’ positive evaluations of these benefits may be substantially weakened when concerns about technological or privacy risks are salient.

Hypothesis H5a.

Perceived Robotaxi risk (PRR) weakens the positive effect of autonomy (PAU) on switching intention (SWI).

Hypothesis H5b.

Perceived Robotaxi risk (PRR) weakens the positive effect of service reliability (PSR) on switching intention (SWI).

4. Method

4.1. Measurement

The measurement instruments in this study consisted of two parts. In the survey design, ride-hailing was defined as mainstream platform-based mobility services represented by economy ride options (e.g., Didi Express, UberX), in which human drivers operate the vehicles and passengers book and pay via mobile platforms. This definition excluded premium chauffeur services or digitalized versions of traditional street-hailing taxis.

The first part of the survey captured demographic information, including gender, age, education, income, driver’s license ownership, and private car ownership, as these variables are closely related to travel mode choice and technology acceptance. The distribution of these variables is summarized in

Table 1. Sample description.

Demographic Information	Category	Frequency	%
Gender	Male	610	50.6
	Female	596	49.4
Age	18–24	366	30.1
	25–34	484	40.1
	35–44	201	16.7
	45–54	118	9.8
	≥55	37	3.1
Education	<High School	179	14.8
	High school or technical school	286	23.7
	Bachelor or Diploma	608	50.4
	≥Master	133	11.0
Monthly Income	<4000 ¥ (<≈ 556$)	182	15.1
	4001–8000 ¥ (≈556–1111$)	511	42.4
	8001–12,000 ¥ (≈1112–1667$)	377	31.3
	≥12,001 ¥ (≥≈1668 USD$)	136	11.3
Driver’s license	Yes	728	60.4
	No	478	39.6
Vehicle ownership	Yes	653	54.1
	No	553	45.9

Note: Income levels (¥) were converted to approximate U.S. dollars (USD) based on the 2025 average exchange rate (1 USD ≈ 7.2 RMB).

, which provides an overview of the respondents’ demographic composition. To complement the demographic profile, additional travel background information was collected, such as frequency of ride-hailing and Robotaxi usage in the past month, proportion of nighttime trips, and negative travel experiences. Figure 2 illustrates the overall mobility frequency, showing that most respondents frequently used both ride-hailing and Robotaxi services, thereby ensuring sufficient real-use experience. Figure 3 depicts the distribution of negative experiences in ride-hailing, where the most frequently reported issues were dangerous driving, detours, and verbal discomfort, highlighting the human-related risks that may push users to consider switching. In contrast, Figure 4 presents negative experiences in Robotaxi, such as recognition failures or inadequate system alerts, reflecting technology-related risks that may act as barriers to adoption. Taken together, Table 1 and Figure 2, Figure 3 and Figure 4 provide a comprehensive characterization of the sample—linking demographic background with behavioral frequency and experiential differences between ride-hailing and Robotaxi—thereby establishing a solid empirical foundation for the subsequent model analysis. To maintain focus and parsimony, the survey concentrated on variables relevant to “travel behavior and migration motives,” and did not include items such as occupation or marital status, which are less relevant to the research theme.

The second part of the survey measured the latent constructs in the research model. All constructs were adapted from established and widely validated scales, with wording adjusted to fit the context of Robotaxi and ride-hailing. In total, seven core constructs were measured, corresponding to the push, pull, and mooring dimensions of the model. All items were assessed using a seven-point Likert scale (1 = strongly disagree, 7 = strongly agree).

To ensure semantic accuracy and cross-cultural validity, a translation–back translation procedure was employed. Two bilingual researchers first translated the original English items into Chinese, and another team independently back-translated them into English for comparison with the original versions. A pilot test was conducted with 30 participants, focusing on the clarity of terminology and contextual relevance. Based on their feedback, final revisions were made to ensure linguistic clarity, contextual fit with the Robotaxi setting, and cultural appropriateness. The measurement items and their sources are listed in Appendix A.

4.2. Sample and Data Collection

These cities were selected based on three criteria. First, they represent the geographical and economic diversity of China’s major urban regions—covering the central (Wuhan, Changsha) and southern coastal (Guangzhou, Shenzhen) areas, which differ in population density, income level, and mobility infrastructure. Second, they are among the earliest and most mature Robotaxi operation zones, with continuous pilot or commercial services since 2022, ensuring that respondents had realistic exposure and usage experience. Third, they collectively capture the typical developmental pattern of Robotaxi services in China, balancing technological advancement, regulatory support, and market acceptance across diverse metropolitan contexts [5,69]. As such, the sample provides a realistic and diversified representation of China’s current stage of Robotaxi deployment, offering an appropriate empirical basis for understanding user migration behavior.

A mixed online–offline approach was employed to ensure that the sample realistically reflected user decisions under migration scenarios and to enhance representativeness. The online survey was conducted through Wenjuanxing, a widely used professional Chinese survey platform that supports forced-response validation and cross-device adaptability, helping to guarantee data accuracy and consistency [70]. The offline survey was administered at frequently used Robotaxi pick-up/drop-off points and major transport hubs in the four cities, where participants accessed the same online questionnaire via QR code. This approach minimized potential mode bias across recruitment channels. All respondents participated voluntarily with informed consent and were required to confirm that they had used both Robotaxi and ride-hailing at least once in the past month before proceeding. To encourage careful and complete responses, participants who successfully completed the survey received a small symbolic incentive (equivalent to USD 1). All data were used solely for academic purposes and strictly adhered to principles of anonymity and privacy protection.

A total of 1600 questionnaires were collected. A unified quality control protocol was applied using multiple exclusion criteria: questionnaires with response times below the 10th percentile (P10) were flagged as invalid, those failing attention-check questions were removed, and incomplete or missing responses were discarded. Additional manual checks were conducted for logical consistency and duplicate submissions to further ensure reliability. After filtering, 394 invalid responses were eliminated, leaving 1206 valid questionnaires, yielding a final effective response rate of 75.4%, which is substantially higher than the average for online survey studies. Data collection lasted approximately six weeks, from mid-March to the end of April 2025.

To assess adequacy of the sample size, the Structural Equation Model sample size calculator was used. With an assumed effect size of 0.3, a statistical power of 0.95, and a significance level of 0.05, the minimum recommended sample size was 247 [71]. Considering the subsequent group analysis, this threshold was adjusted to 494. The final sample of 1206 respondents comfortably exceeded these requirements, ensuring robust statistical power for the analyses.

4.3. Common Method Bias

To further assess common method bias (CMB), both Harman’s single-factor test and the full collinearity VIF approach were applied. The single-factor test indicated that the first factor accounted for 30.98% of the total variance, below the recommended threshold of 40% [72]. In addition, all inner and outer VIF values were below 3.3 (

Table 2. Scale items and convergent validity analysis.

Constructs	Codes	VIF (OM)	VIF (IM)	λ	α	CR	AVE
Social Anxiety	SAN1	1.828	1.324	0.833	0.838	0.892	0.673
(SAN)	SAN2	1.891		0.829
	SAN3	1.812		0.817
	SAN4	1.741		0.802
Human-Induced	HII1	1.964	1.333	0.839	0.853	0.901	0.694
Insecurity	HII2	2.037		0.851
(HII)	HII3	1.879		0.818
	HII4	1.876		0.825
Perceived Autonomy	PAU1	1.897	1.521	0.825	0.854	0.901	0.695
(PAU)	PAU2	1.937		0.841
	PAU3	1.988		0.834
	PAU4	1.926		0.835
Perceived	PSR1	1.888	1.267	0.826	0.845	0.896	0.682
Service Reliability	PSR2	1.878		0.834
(PSR)	PSR3	1.831		0.812
	PSR4	1.852		0.831
Habit	HAB1	1.848	1.186	0.823	0.845	0.896	0.683
(HAB)	HAB2	1.894		0.824
	HAB3	1.902		0.838
	HAB4	1.829		0.819
Perceived	PRR1	2.084	1.464	0.853	0.854	0.901	0.695
Robotaxi Risk	PRR2	1.986		0.841
(PRR)	PRR3	1.865		0.822
	PRR4	1.854		0.819
Switching Intention	SWI1	1.697		0.847	0.790	0.877	0.705
(SWI)	SWI2	1.685		0.841
	SWI3	1.614		0.830

), suggesting that multicollinearity was not a concern and that CMB was unlikely to pose a threat to the validity of the results [73].

5. Results

This study employed partial least squares structural equation modeling (PLS-SEM) to test the proposed model. PLS-SEM is particularly suitable for complex models involving multiple latent variables, moderating effects, and multi-group comparisons. Unlike covariance-based SEM, which emphasizes theory confirmation, PLS-SEM prioritizes exploratory analysis and theory development, making it well-suited for the objectives of this study. All analyses were conducted using SmartPLS 3.29.

5.1. Measurement Model

To ensure the reliability and validity of the measurement model, several assessments were performed. As shown in Table 2, all item loadings exceeded the threshold of 0.70, while Cronbach’s α and composite reliability (CR) values were all above 0.70, confirming internal consistency. Average variance extracted (AVE) values were greater than 0.50, indicating satisfactory convergent validity [74].

Discriminant validity was examined using the Fornell–Larcker criterion (

Table 3. Fornell-Larcker criterion.

	HAB	HII	PSR	SAN	PAU	PRR	SWI
HAB	0.826
HII	−0.222	0.833
PSR	−0.258	0.293	0.826
SAN	−0.284	0.332	0.326	0.820
PAU	−0.305	0.401	0.379	0.377	0.834
PRR	0.301	−0.404	−0.308	−0.368	−0.468	0.834
SWI	−0.290	0.414	0.365	0.412	0.433	−0.541	0.839

) and the heterotrait–monotrait ratio (HTMT,

Table 4. HTMT criterion.

	HAB	HII	PSR	SAN	PAU	PRR	SWI
HAB
HII	0.261
PSR	0.303	0.344
SAN	0.336	0.393	0.386
PAU	0.360	0.471	0.446	0.446
PRR	0.353	0.473	0.361	0.434	0.548
SWI	0.354	0.503	0.446	0.505	0.526	0.658

). The square root of each construct’s AVE was higher than its inter-construct correlations [75], and all HTMT values were below 0.85 [76], demonstrating robust convergent and discriminant validity for the model.

5.2. Model Fit, Explanatory Power, and Predictive Power

The model fit indices indicated satisfactory results. The standardized root mean square residual (SRMR) was 0.039, below the recommended threshold of 0.08 [77], confirming good model fit (

Table 5. Model’s fit.

	Fit Indices
SRMR	0.039
d_ULS	0.569
d_G	0.232
NFI	0.889

). The d_ULS (0.569) and d_G (0.232) values were within acceptable ranges, while the normed fit index (NFI) was 0.889, close to the recommended cut-off of 0.90, further supporting model adequacy.

In terms of explanatory power, the R² for switching intention (SWI) was 0.402, with an adjusted R² of 0.399, indicating that the model explained approximately 40% of the variance in the dependent variable—a moderate level of explanatory power [78]. Moreover, the Q² value for SWI was 0.279, which is greater than zero, confirming that the model possessed predictive relevance for the endogenous construct (

Table 6. Explanatory power and predictive power.

	Q2	R2	Adjusted R2
SWI	0.279	0.402	0.399

). Taken together, the findings demonstrate that the model achieved satisfactory performance in terms of fit, explanatory power, and predictive power [79].

5.3. Direct Effect Test

The results of the structural model are presented in

Table 7. Hypothesis testing results.

Hypotheses	Paths	β	T	P	0.025	0.975	Results
H1a	SAN → SWI	0.143	4.922	0.000	0.087	0.200	√
H1b	HII → SWI	0.114	4.104	0.000	0.062	0.170	√
H2a	PAU → SWI	0.112	3.387	0.001	0.048	0.178	√
H2b	PSR → SWI	0.115	3.433	0.001	0.051	0.182	√
H3a	HAB → SWI	−0.060	2.435	0.015	−0.110	−0.014	√
H3b	PRR → SWI	−0.362	10.482	0.000	−0.428	−0.293	√
H4a	HAB × SAN → SWI	0.039	1.249	0.212	−0.023	0.101	×
H4b	HAB × HII → SWI	0.070	2.593	0.010	0.017	0.123	×
H5a	PRR × PAU → SWI	−0.131	3.619	0.000	−0.200	−0.058	√
H5b	PRR ×PSR → SWI	−0.120	3.468	0.001	−0.186	−0.051	√

. Social anxiety exerted a significant positive effect on switching intention (β = 0.143, p < 0.001), indicating that heightened social discomfort motivates users to seek alternative travel solutions such as Robotaxi, thereby supporting H1a. Similarly, human-induced insecurity (β = 0.114, p < 0.001) also showed a positive influence on switching intention, suggesting that concerns about driver reliability increase users’ acceptance of autonomous driving services, supporting H1b.

In addition, perceived autonomy (β = 0.112, p < 0.01) and perceived service reliability (β = 0.115, p < 0.01) both had significant positive effects on switching intention. These findings highlight the importance of enhancing users’ sense of control and trust in service performance in facilitating behavioral change, thereby supporting H2a and H2b.

By contrast, habit (β = –0.060, p < 0.05) negatively predicted switching intention, showing that entrenched behavioral patterns remain a significant barrier to adopting Robotaxi services, thereby supporting H3a. Perceived Robotaxi risk (β = –0.362, p < 0.001) demonstrated the strongest negative effect, confirming that safety and performance concerns substantially hinder users’ willingness to switch, thereby supporting H3b.

5.4. Moderating Effect

The results of the moderating effect tests are summarized in Table 7 and visualized in Figure 5, Figure 6, Figure 7 and Figure 8. As illustrated in Figure 5, the slopes for different levels of habit are nearly parallel, indicating that habitual tendencies do not significantly alter the impact of social anxiety on switching intention. This suggests that regardless of their prior ride-hailing experience, users with high social anxiety remain motivated to avoid interpersonal discomfort by choosing Robotaxi.

In contrast, habit significantly moderated the path from human-induced insecurity to switching intention (H4b) (β = 0.070, p = 0.010). However, the direction of the effect was opposite to the hypothesized negative moderation. Specifically, as illustrated in Figure 6, stronger habitual tendencies unexpectedly amplified the positive relationship between human-induced insecurity and switching intention, rather than weakening it. This suggests that users with stronger ride-hailing habits become more aware of driver-related insecurity and are therefore more inclined to switch to Robotaxi as a safer, more predictable alternative. Thus, H4b was not supported.

Regarding perceived Robotaxi risk, Figure 7 and Figure 8 jointly display its moderating role on the pull factors. The effect of autonomy on switching intention (Figure 7) becomes weaker when perceived risk is high, while a similar dampening effect is observed for service reliability (Figure 8). These results imply that even when Robotaxi offer technological advantages, high perceived risk can suppress users’ positive evaluations and slow their behavioral migration. Finally, integrates these findings into the full structural model (Figure 9), illustrating how push and pull factors jointly promote migration while habit and perceived risk act as contextual boundaries that shape users’ switching behavior.

5.5. Multi-Group Analysis

Before conducting the multi-group analysis (MGA), it was necessary to test for measurement invariance. Gender was selected as the grouping variable because, compared with other demographic factors such as age or education level, previous studies have consistently shown that gender plays a more decisive role in shaping users’ risk perception, safety sensitivity, and technology evaluation in autonomous mobility contexts. These aspects are also central to the psychological mechanisms examined in this study [22,80,81]. Following prior research and methodological guidelines [81,82], the MICOM procedure was applied to examine configural invariance, compositional invariance, and equality of means and variances. Establishing configural and compositional invariance is sufficient for proceeding with MGA. As shown in

Table 8. Assessment of Measurement Invariance.

Group	Con.	Conf. Inv.	Comp. Inv.		PMI Est.	Equal Mean		Equal Var		FMI Est.
			C = 1	CI.		Diff.	CI.	Diff.	CI.
Male	SAN	Yes	0.998	[0.999; 0.998]	Yes	−0.038	[−0.110; 0.109]	−0.175	[−0.167; 0.159]	NO
VS	HII	Yes	1.000	[0.999; 0.999]	Yes	−0.188	[−0.108; 0.107]	0.109	[−0.112; 0.121]	NO
Female	PAU	Yes	0.999	[1.000; 0.999]	Yes	−0.158	[−0.131; 0.121]	−0.077	[−0.115; 0.123]	NO
	PSR	Yes	0.999	[0.999; 0.998]	Yes	−0.197	[−0.125; 0.122]	0.028	[−0.126; 0.134]	NO
	HAB	Yes	0.999	[0.999; 0.996]	Yes	0.122	[−0.111; 0.120]	−0.091	[−0.123; 0.121]	NO
	PRR	Yes	0.999	[1.000; 0.999]	Yes	0.264	[−0.116; 0.118]	0.161	[−0.120; 0.110]	NO
	SWI	Yes	1.000	[1.000; 0.999]	Yes	−0.227	[−0.114; 0.109]	0.283	[−0.154; 0.163]	NO

Note: Con.: Construct; Conf. Inv.: Configural Invariance; Comp. Inv.: Compositional Invariance; CI: Confidence Interval; PMI Est.: Partial Measurement Invariance Established; Equal Mean: Equal Mean Value; Diff.: Differences; Equal Var: Equal Variances; FMI Est.: Full Measurement Invariance Established.

, all constructs satisfied both configural and compositional invariance, achieving partial measurement invariance. This indicates that the measurement model is comparable across gender groups, allowing for a meaningful examination of structural path differences. This result confirms that any subsequent group differences in the structural model can be attributed to true behavioral variation rather than to measurement bias.

The MGA results are presented in

Table 9. Fornell-Larcker criterion. Multigroup analysis results.

Group	Relationship	β0	β1	Coefficient Difference
Male (0)	SAN → SWI	0.056	0.248 ***	−0.192 **
VS	HII → SWI	0.085 *	0.193 ***	−0.108 *
Female (1)	PAU → SWI	0.103 *	0.110 **	−0.007
	PSR → SWI	0.168 ***	0.023	0.144 *
	HAB → SWI	−0.071	−0.035	−0.036
	PRR → SWI	−0.261 ***	−0.418 ***	0.157 *

Note: * p < 0.05; ** p < 0.01; *** p < 0.001.

. The effect of social anxiety on switching intention (SAN → SWI) was significantly stronger among females than males (difference = −0.192, p < 0.01), suggesting that women are more likely to choose Robotaxi to alleviate social discomfort. Similarly, the path from human-induced insecurity to switching intention (HII → SWI) also showed gender differences (difference = −0.108, p < 0.05), with women being more sensitive to driver-related safety concerns. In contrast, men exhibited a stronger positive effect of perceived service reliability on switching intention (PSR → SWI) (difference = 0.144, p < 0.05), indicating that male users place greater importance on system reliability and efficiency.

Notably, in the case of perceived Robotaxi risk (PRR → SWI), both groups showed significant negative effects, but the effect was stronger for females (β = −0.418, p < 0.001) than for males (β = −0.261, p < 0.001), highlighting women’s higher sensitivity to technological uncertainty. Meanwhile, no significant gender differences were found in the effects of perceived autonomy (PAU → SWI) or habit (HAB → SWI), suggesting that these factors operate similarly across male and female groups. Overall, these results reveal that gender differences mainly arise from risk-related and performance-related evaluations, while autonomy and habit remain consistent across users.

6. Discussion

6.1. Theoretical Contributions

This study offers several theoretical contributions to understanding user adoption of autonomous taxis.

First, unlike most prior studies on autonomous driving that have applied TAM, UTAUT, or trust-based models and treated adoption as an isolated decision [6,7]. Such assumptions overlook the fact that in reality, users already have mature alternatives—namely, ride-hailing—and their decision logic is essentially about whether to migrate rather than whether to adopt. By applying the PPM framework, this study analyzes Robotaxi adoption as a competitive migration process vis-à-vis ride-hailing, uncovering the dynamic interplay of push, pull, and mooring forces. This perspective extends the application of PPM to high-risk innovations and shifts autonomous driving research from a “single-point adoption logic” toward a “competition-and-substitution logic,” thereby enhancing its explanatory power.

Second, prior research has typically regarded perceived risk as a unidimensional factor that simply reduces adoption intention [14,83]. In migration contexts, however, risk sources are not homogeneous. This study distinguishes between human-related risks and technology-related risks and reveals their functional differences: human-related risks act as push forces driving users away from ride-hailing, while technology-related risks serve as mooring forces that hinder switching to Robotaxi. This finding addresses the tendency of prior research to generalize risk and provides a more functional explanatory framework, showing that risk is not always a barrier to adoption but may either propel or inhibit migration depending on the context.

Third, by examining moderating effects, this study uncovers boundary conditions and complexities in the formation of switching intentions. Some findings align with expectations—for instance, perceived risk attenuated the positive effects of autonomy and service reliability. Yet others contradicted the hypothesized direction. A theoretically significant and unexpected result was that habit, rather than weakening the influence of human-induced insecurity, actually strengthened its positive effect on switching intention. Habit may play this strengthening role because, consistent with the risk amplification [84] and the theory of cognitive dissonance [85] perspectives, repeated exposure to driver-related uncertainty reinforces users’ sensitivity to human insecurity. When such insecurity accumulates within habitual ride-hailing experiences, it creates psychological tension and motivates users to seek a driverless alternative to restore perceived safety. This suggests that habit is not merely a barrier in migration research but may act as a risk amplifier in high-risk contexts, intensifying users’ consideration of migration. This revises the narrow view of habit as solely path dependence and resistance to change, and it resonates with the broader definition of mooring factors in the original PPM framework, which recognized that mooring can both constrain and facilitate migration. Theoretically, this underscores the importance of considering duality and contextual dependency in studying autonomous driving adoption.

Finally, extending prior research that has noted gender differences in mobility preferences—where women emphasize safety and privacy while men value efficiency and autonomy—this study systematically validates these differences in the context of Robotaxi migration [19,20]. Consistent with previous findings, women exhibit stronger sensitivity to social anxiety, human-induced risk, and autonomous-driving risk, whereas men rely more heavily on service reliability. This not only consolidates scattered evidence of gendered differences in consumer behavior and mobility but also establishes, for the first time, a systematic comparative framework for gender differences in Robotaxi migration. In doing so, the study extends the boundaries of individual difference theory in the context of high-risk innovation adoption.

6.2. Practical Implications

Beyond its theoretical value, this study provides several practical insights for the promotion and implementation of Robotaxi services, relevant to service providers, policymakers, and system designers.

First, the findings highlight the necessity of gender-sensitive promotion strategies. This study shows that female users are more strongly affected by social anxiety, human-induced risks, and technological risks, while male users rely more heavily on perceptions of service reliability. Accordingly, service providers should adopt differentiated strategies in marketing and communication [86,87]. For female users, emphasis should be placed on transparent safety assurance mechanisms—such as real-time monitoring during trips, emergency alerts in abnormal situations, and responsive customer support—as well as robust privacy protection measures to mitigate risk perception. For male users, providers should highlight system stability and service consistency, for instance by showcasing Robotaxi’s adherence to traffic rules, smooth driving, and operational reliability. Such targeted, group-specific strategies can enhance market penetration and avoid the inefficiencies of a “one-size-fits-all” approach.

Second, the unexpected role of habit offers new directions for Robotaxi promotion. Traditional perspectives often treat habit as a barrier to migration, arguing that users are locked into existing routines and reluctant to switch [62]. However, this study finds that when frequent ride-hailing users perceive insecurity from drivers, habit does not diminish this effect but instead amplifies its impact on switching intention. This suggests that service providers should adopt a dual-path strategy: on the one hand, reduce the lock-in effect of old habits through free trial rides, subsidies, or collaborations with existing mobility platforms; on the other hand, leverage frequent users’ heightened sensitivity to human-induced insecurity by emphasizing Robotaxi’s superior safety and consistency. By simultaneously weakening old habits and capitalizing on new ones, service providers may significantly accelerate market penetration.

Finally, the study underscores the importance of social and psychological drivers in promoting Robotaxi adoption. Results indicate that social anxiety and human-induced risks are key push forces motivating users to switch [88]. Service providers should highlight Robotaxi’s unique value propositions—such as reducing driver–passenger social pressure, avoiding human driving errors, and offering a quiet, autonomous travel space—to establish clear market differentiation. At the same time, policymakers and urban authorities should strengthen transparency and external safeguards at the institutional level. These may include clarifying liability in accidents, establishing unified safety and operational standards, and implementing robust data security and privacy protection frameworks. Such measures not only reduce perceived risks, particularly among higher-income groups, but also provide institutional legitimacy and build public trust, both of which are essential for large-scale Robotaxi deployment.

6.3. Limitations and Future Research

First, the data were collected in actual operational contexts in Chinese cities. While this enhances external validity, it remains uncertain whether the findings can be generalized to other cultural or regional settings. Future research could conduct cross-context comparisons across cities or countries to test the robustness and universality of migration logics and risk mechanisms.

Second, the study primarily focused on factors such as perceived risk, habit, social anxiety, autonomy, and service reliability. Although these variables provide strong explanatory power, other potentially important factors—such as environmental attitudes, brand image, or psychological ownership—were not included. Future research could integrate these constructs to build a more comprehensive model of migration mechanisms and deepen understanding of users’ psychological and behavioral processes.

Third, the study relied on cross-sectional data, which limits the ability to establish causal relationships or capture the dynamic evolution of user attitudes. Longitudinal tracking or experimental designs could be employed in future work to more precisely capture the process through which users move from initial contact with Robotaxi to the formation of stable usage intentions, thereby incorporating the temporal dimension of migration behavior.

Finally, although this study examined moderating effects and gender differences, it did not fully capture the interactions between institutional, technological, and social factors. Future research could adopt more complex, multi-level frameworks that combine institutional environments, technical performance, and social–psychological mechanisms. Moreover, as Robotaxi fleets expand and gradually replace traditional taxis, age-related disparities in adoption may emerge. Older passengers, who often exhibit lower technology familiarity and higher risk sensitivity, may become less willing to use driverless services and instead rely more on public transportation. Investigating this potential substitution trend would contribute to the development of inclusive mobility strategies and prevent the marginalization of elderly users during the transition toward autonomous mobility systems.

7. Conclusions

This study reframes Robotaxi adoption as a migration process within competing mobility services, rather than a vacuum-based act of technology acceptance. Using the Push–Pull–Mooring framework, we demonstrate that switching intentions are shaped by both human-induced and technology-related factors. Social anxiety and insecurity stemming from human drivers act as push forces, while perceived autonomy and service reliability act as pull forces. Conversely, entrenched ride-hailing habits and perceived technological risks function as mooring forces that inhibit switching. Notably, this study distinguishes between human- and technology-related risks, revealing their opposite roles in shaping behavioral outcomes.

Our findings further highlight heterogeneity in user responses. Gender-based multi-group analysis shows that women are more sensitive to safety and privacy concerns, while men emphasize efficiency and reliability. Moreover, habits, traditionally considered a barrier, were found to amplify the effect of insecurity, suggesting that negative experiences in ride-hailing may accelerate migration when combined with entrenched usage patterns. These results underscore the complex interplay between human–machine interaction, psychological mechanisms, and user differences in the context of autonomous driving.

Theoretically, this study contributes to the literature by extending the Push–Pull–Mooring framework to the emerging field of autonomous mobility and by revealing how differentiated sources of risk (human-induced vs. technology-related) jointly shape migration intentions. This theoretical refinement deepens understanding of behavioral migration in high-risk innovation contexts and broadens the explanatory scope of the PPM framework beyond traditional technology-adoption settings, thereby contributing new conceptual insights to both migration and risk theories. Practically, it offers actionable insights for Robotaxi operators and policymakers, such as designing gender-sensitive strategies, emphasizing transparent safety communication, and addressing user heterogeneity to foster broader acceptance. Furthermore, the methodological procedure developed in this study is highly replicable. The integration of the Push–Pull–Mooring framework with multigroup PLS-SEM provides a transparent analytical process that can be easily reproduced in other regional or technological contexts to examine user migration behaviors under varying conditions.

Beyond these theoretical and practical implications, this study also recognizes certain contextual boundaries that warrant further examination. Although this study was conducted within the Chinese context, the core logic of the Push–Pull–Mooring framework reflects universal migration mechanisms of dissatisfaction, attraction, and inertia. Therefore, the model is expected to remain conceptually applicable in other cultural settings, although the relative strengths of these forces may vary across contexts. Looking ahead, future research could conduct cross-cultural and longitudinal investigations to validate these contextual variations and capture the evolving dynamics of autonomous mobility adoption. In doing so, scholars can better understand how human–machine interaction, behavioral differences, and risk perceptions jointly shape the sustainable transition from conventional ride-hailing to autonomous transport solutions. In addition, future studies could explore other psychological and contextual drivers of migration behavior—such as trust, perceived control, or environmental attitudes—examine user differences across age and experience groups, and extend the PPM framework to other shared or automated mobility modes beyond Robotaxi.