Introduction

Background

Autonomous vehicles (AVs) are emerging as a highly promising advancement in transportation. These vehicles are currently undergoing trials in multiple countries, including eight ongoing projects in Australia (KPMG International 2020; Austroads 2021). Research has shown that AVs have the potential to revolutionise the way we travel and significantly impact urban land use development (Soteropoulos et al. 2019; Yigitcanlar et al. 2019).

Land use and transportation share a complex and mutually influential relationship, which can be theoretically explained through the transport land use feedback cycle (Bertolini 2012). This feedback cycle depicts a reciprocal interaction between land use and transport, with activity and accessibility as interconnected components. The emergence of AVs is expected to disrupt the existing system of transportation and land use, as illustrated in Fig. 1. The use of AVs has a direct impact on transportation, influencing factors such as travel costs and values of travel time. These changes subsequently affect the accessibility experience of individuals. Accessibility, defined as the ease of reaching activity locations (Handy and Niemeier 1997), has different effects over different time scales. In the short term, it influences discretionary location decisions, while in the long term, it affects long-term location decisions, such as residential dwellings, by changing the attractiveness of certain areas. Both short-term and long-term location decisions alter individuals’ activity patterns, resulting in new conditions on the transport network. Bertonili’s theory has provided an overview of interactions between key components of urban environments, establishing a theoretical foundation for analysing and predicting the potential impact of AVs.

Fig. 1
figure 1

Effects of AVs on the land use transport feedback cycle (Adapted from Bertolini (2012))

Full size image

There are still many uncertainties regarding the effect of AVs on location decisions. Existing research primarily focuses on long-term location decisions, particularly residential location choices (Thakur et al. 2016; Zhang and Guhathakurta 2021). However, the impact of AVs on discretionary location choices, also known as destination choices, is rarely discussed separately and mostly inferred through metrics such as vehicle miles travelled (VMT) in travel demand analyses (Childress et al. 2015; Auld et al. 2017; Kröger et al. 2019). While most studies predict an increase in VMT with the use of AVs, it remains unclear whether this increase is attributed to changes in discretionary location choices. As mentioned by Milakis et al. (2017), discretionary location choices are expected to experience first-order implications of AVs, potentially occurring in the near future. Therefore, it is imperative to explore the interaction between AVs and the location choice for discretionary activities to gain a better understanding of the immediate impact of AVs.

Related work

Numerous studies have examined the capabilities of AVs through surveys, field experiments, and analytical approaches (Bagloee et al. 2016; Faisal et al. 2019; Harb et al. 2021). One important feature of AVs is their self-driving operation, which allows drivers to engage in other productive activities during travel and perceive in-vehicle travel time less negatively than in traditional cars. This effect can be quantified as a decrease in the value of travel time (VOTT). A reduced VOTT may lead individuals to travel longer distances for a wider variety of products or fresher goods in the case of discretionary activities such as grocery shop**. Regarding long-term location decisions, individuals may relocate to outer suburbs with relatively lower housing prices despite the trade-off of longer commutes. Another anticipated capability is the potential reduction in operational costs with the shift towards electric vehicles (Davidson and Spinoulas 2015). Additionally, some researchers predict a fully shared AV scheme where no private vehicle is used (Heilig et al. 2017; Zhang and Guhathakurta 2021). By eliminating expenses associated with purchasing and maintaining vehicles, the overall cost of a vehicle is expected to decrease compared to traditional car ownership. Lastly, AVs have the potential to increase road capacity by enabling vehicle communication and collaboration, resulting in shorter headway between vehicles and increased traffic volume. However, this outcome is more likely to be achieved when AV market penetration is high (Carrone et al. 2021).

Changes in transportation have a significant impact on accessibility. Accessibility is a comprehensive measure to describe the net effect between land use, transport, individual characteristics and temporal dimension (Geurs and Wee 2004). Various types of accessibility measures exist, each emphasising different aspects of access experiences. Place-based accessibility measures focus on the spatial interaction between an origin and all other accessible locations or services, with results largely determined by existing land use patterns and transport conditions (Miller 2005). These measures are commonly utilised in land use-transportation interaction (LUTI) models, such as UrbanSim (Waddell 2002) and SILO (Moeckel 2017), to support predictions of long-term location choices. The common measures of place-based accessibility include gravity-based measures and utility-based measures. In contrast, person-based accessibility measures adopt an individual-centered perspective and consider the spatio-temporal constraints on an individual’s movements. It is argued that individuals often have fixed activities, such as going to work, which are difficult to relocate or reschedule in their daily routines. These fixed activities limit an individual’s ability to access places for discretionary activities. Therefore, person-based accessibility measures focus on the land use and transport supplies that are valid to individuals within their specific spatio-temporal limitations (Miller 1991; Kwan 1998). However, the implementation of person-based accessibility measures is scarce due to the difficulty of accessing the large amount of diary data needed for their calculation (Geurs and Wee 2004).

Several studies have estimated the potential impact of AVs on individuals’ activity patterns. Childress et al. (2015) employed an activity-based travel demand model to test various AV deployment scenarios. Their findings suggest that with a 30% increase in road capacity on freeways and major arterial roads, a 35% reduction in VOTT, and a halved parking cost, work trip lengths for Puget Sound residents may increase by up to 61%, while school trip lengths may increase by up to 15%. This indicates that individuals may be more tolerant of longer commuting distances to workplaces and schools. In a quasi-real case study conducted in Delft, Correia and Arem (2016) studied the impact of AVs on activities during a typical working day. They concluded that a 50% reduction in VOTT can lead to a 49% increase in VMT, along with a 24% increase in driving time per trip. Kröger et al. (2019) compared different AV market shares in Germany and the USA using an aspatial travel demand model. The authors categorised education and work trips as commute trips and considered all other trips as non-commute trips. Assuming a 25% lower VOTT, simulation results show that a low market share causes a 3.4% increase in vehicle kilometers travelled (VKT), while a high market share results in an 8.6% increase in VKT. Auld et al. (2017) utilised an activity- and agent-based model called POLARIS to examine the effect of various VOTT reductions, road capacities, and market shares on travel demand. They found that all these factors contribute to a significant increase in VMT. However, the authors believed that person-based accessibility consideration might impose constraints on spatial movements, limiting the extent of increases in trip lengths.

Research focus

We identified two knowledge gaps from the literature review. The first gap pertains to the limited attention given to the impact of AVs on discretionary location choices. Existing studies tend to model trips for all purposes, including both commutes (education and work trips) and discretionary activity trips. However, discretionary activity trips have different temporal dynamics, as they can be more easily and quickly altered. In contrast, changes in job and school locations typically occur less frequently for most individuals. Therefore, it is important to address these temporal dynamics and analyse different types of behaviour separately when assessing the changes in location choices due to AVs. The second knowledge gap is the lack of consideration for accessibility when estimating the impact on activity patterns. Existing studies overlook spatio-temporal constraints on individuals’ movements, potentially resulting in over-optimistic estimates of increases in trip lengths and VMT. To bridge these gaps, we conducted a case study of Clayton, Australia, to assess the impact of AVs on discretionary location choices using an agent-based integrated simulation framework. This framework incorporates an activity-based travel demand model that considers person-based accessibility constraints. Various scenarios involving different reductions in VOTT and travel costs were analysed. Considering that the initial stage of AV deployment might be mainly private AVs (Davidson and Spinoulas 2015), the case study is limited to this context and assumes AVs are not allowed to drive unoccupied. This research simulates a near-future scenario that captures the effect of AVs during their early stage, aiming to provide timely insights for policymakers and planners to prepare for the emergence of AVs.

The rest of the paper has the following structure. “Proposed simulation framework” section introduces the simulation framework and describes the details of framework components. “Case study of Clayton” section demonstrates the case application located in Clayton, including the details of simulations and a discussion of findings. “Conclusion” section provides the conclusion of the paper. Lastly, “Study limitations and future work” section discusses the limitations of this work and identifies opportunities for future research.

Proposed simulation framework

Framework overview

The proposed simulation framework used in this study is an integration of an activity-based travel demand model, mobiTopp (Mallig et al. 2013), with a multi-agent transport simulator, MATSim (Axhausen et al. 2016). MobiTopp generates a synthetic population with their corresponding demographics and one-day activity schedules. Details of activity schedules are complemented through discretionary location choice and mode choice modelling, with an accessibility-constraining process influencing the choice set of discretionary locations. The output of the travel demand model is a detailed one-day activity schedule for the synthetic population. The activity-based travel demand is then simulated in MATSim. As the performance of motorised traffic reaches convergence in traffic assignment, zonal travel times are computed and fed back to the travel demand model, thereby impacting accessibility constraints and choice behaviour. The simulation continues until predefined stop** criteria are satisfied. Further details about each framework component are elaborated in subsequent sections. Figure 2 demonstrates the data flow in the simulation framework.

Fig. 2
figure 2

Data flow in the simulation framework

Full size image

Travel demand with accessibility constraints

The demand generation process within the travel demand model consists of two phases. The first phase is the generation of population and activity patterns. The second phase is a series of choice modelling that assigns locations and travel modes to the activity patterns. During the first phase, the population generation model uses locational attributes and travel surveys to construct a virtual built environment with corresponding demographics. The model applies a hierarchical iterative proportional fitting procedure (Müller and Axhausen 2011) to generate a synthetic population. Each simulated person has an activity pattern generated by the activity generation model. This model uses multinomial logistic regression to sequentially model travel decisions, such as the main activity type and the number of tours in a day (Hilgert et al. 2017). The outcome of the activity generation model is a pattern of activities in one day with only preliminary information. This pattern describes each activity’s type, planned start time and duration, with no location and mode choice details.

Fig. 3
figure 3

Demonstration of the accessibility constraint

Full size image

After obtaining the preliminary activity pattern, the second phase of demand generation involves location and mode choice modelling. The preliminary activity pattern includes three activity categories: home, fixed and discretionary. The fixed activity category represents mandatory activity, such as work and school, while the discretionary activity category includes subcategories such as shop** and leisure. The location choice model is applied only to determine the locations of discretionary activities, while the locations of home and fixed activities are extracted from surveys and remain unchanged throughout the demand generation process. Prior to location choice modelling, an accessibility-constraining process is introduced. In this process, it is assumed that fixed activities have a fixed starting time, and simulated people must avoid being late for all fixed activities. Therefore, the location of the previous activity should not impede the arrival time at the location of fixed activities. To avoid unwanted delays, the accessibility constraint evaluates the person-based accessibility of individuals before they commence discretionary activities and delimits the accessible locations. Figure 3 demonstrates the accessibility constraint using a snippet of an activity pattern. This snippet involves five activities and four trips between Home Activity 1 (HA1) and Discretionary Activity 2 (DA2). When simulated individuals finish Fixed Activity 1 (FA1), they begin to choose a location for Discretionary Activity (DA1), knowing they should arrive at Fixed Activity 2 (FA2) by a required time. Consequently, the location choice set of DA1 is bounded by a time budget composed of access from FA1 to DA1, duration at DA1 and egress from DA1 to FA2. This time budget can be illustrated as a spatio-temporal constraint on a person’s travel path area (Kwan 1998). This accessibility constraint is applied to discretionary activities only when scheduled before a fixed activity. For example, if there is no fixed activity after DA2, the choice set of DA2 is not constrained.

The location choice model is implemented using a multinomial logit model based on the theory of random utility in which individuals always choose a location that maximises their utility (Ben-Akiva and Lerman 1985). The modelling is made at the level of traffic analysis zones (TAZs), and the exact location is randomly selected from the chosen TAZ. The utility of Individual n choosing destination Zone i is specified as:

$$\begin{aligned} \begin{aligned} {V_{in}}&= f\left( A_{i}, T_{ij}, \frac{C_{ij}}{I_{n}}, \epsilon _{ij}\right), \\ \end{aligned} \end{aligned}$$
(1)

where \(A_{i}\) is the locational attractiveness of Zone i, \(T_{ij}\) is the vector of travel time impedance between destination Zone i and origin Zone j, \(C_{ij}\) is the vector of travel costs between Zone i and Zone j derived from survey-based car running costs (RACV 2022), \(I_{n}\) is the monthly income of Individual n, and \(\epsilon _{ij}\) is the unobserved random portion of utility. Readers can refer to the mobiTopp model (Mallig and Vortisch 2. Inside MATSim, each simulated person is referred to as an agent in the network. During the initial iteration, all activity schedules are loaded into the network, and each agent is assigned an initial travel plan (for example, route or departure time) to execute their activities. Subsequently, all plans are run simultaneously, and each agent’s plan is evaluated with a score that represents the net utility. In the replanning phase, agents may change their plans for the next iteration to gain a higher utility score. This iterative loop between the mobsim and replanning phase is a co-evolutionary process that optimises individual scores of each agent’s plan, ensuring that each agent has the best plan and cannot obtain a higher score. Once the scores of all agents’ plans have been optimised, the traffic assignment reaches convergence.

Hourly travel time matrices of the network are computed once traffic assignment is converged, and these matrices are then used to update the travel time between each TAZ in the travel demand model. The updated travel time data is utilised in the second phase of demand generation within mobiTopp. This iterative process continues until the stop** criteria are met, indicating that the integration between the travel demand model and the dynamic traffic assignment model has reached equilibrium. However, it is important to note that the concept of this equilibrium is not fully understood in current research (Juri et al. 2016). Typically, there are two types of approaches used as the stop** criterion. One is the fixed-point approach that seeks stable changes in origin–destination trips or travel-time skims in successive iterations (Lin et al. 2008). Another approach is to define a fixed number of iterations (Hao et al. 2010). The latter approach is often more practical for large-scale simulations, as it provides a manageable running time. Once the stop** criteria are satisfied, the simulation concludes and the output is activity-based travel diaries for the synthetic population.

Case study of Clayton

Study area and data sources

In this case study, we applied the simulation framework to investigate the impact of AVs on discretionary location choices in Clayton, a populated residential suburb in Southeast Melbourne. The study area encompasses a 520 km2 region, with Clayton at the center and extending up to 12 km in all directions, as illustrated in Fig. 4. Our modelling focused on agents residing in Clayton and their daily activity-based travel demand across the entire study area. To facilitate a detailed analysis, the study area is divided into 642 TAZs based on the boundary of destination zones defined by the Australian Bureau of Statistics (2021). The use of small zones enables a high-resolution study, allowing for the identification of nuanced differences in discretionary activity location choices. The Victorian Integrated Survey of Travel and Activity (VISTA)Footnote 1 reveals that the 90th percentile trip length of Clayton residents engaging in shop** and leisure activities is 12 km. Hence, the study area is designed to cover this range.

The VISTA data provided demographic information and travel behaviour data from over 27,000 households and 69,000 individuals surveyed between 2012 and 2018. Locational attributes and transport networks were obtained from OpenStreetMap, an open-source worldwide geographic database. The transport network was cleaned and validated using the PTV Visum model (PTV AG 2022). Various matrices, including inter-zonal travel times and costs for each travel mode, were generated using the validated network in Visum. These datasets enable mobiTopp to construct a virtual built environment and assign demographics to households during the first phase of demand generation.

Several modelling parameters were calibrated to fit the case study. The activity generation model was not applied in this case study as a full-day activity pattern of the simulated population can be extracted from the VISTA data. The location choice and mode choice models were calibrated using the VISTA data, ensuring they captured the preferences and behaviour of travellers in Melbourne. Detailed calibration results are provided in the “Appendix”. Default parameter values were applied to MATSim, which is primarily used as a traffic assigner in this study.

Fig. 4
figure 4

Study area

Full size image

Scenario setup and assumptions

A scenario-based experimental approach was used in this case study. We simulated all activities from the synthetic Clayton population over 24 h, which accounts for about 15,664 agents and 66,124 trips within the study area. To further construct a reasonable simulation environment, we combine the insights drawn from Davidson and Spinoulas (2015) and present a summary of scenarios in Table 1. We also demonstrate how person-based accessibility constraints affect the behavioural analysis of using AVs. The following sections elaborate on design considerations and assumptions.

Table 1 Summary of scenarios
Full size table
  1. 1.

    Value of travel time The AVs in the case study possess level 4 vehicle autonomy, enabling them to perform all self-driving operations. This function allows for a reduction in VOTT, and the case study explores two levels of VOTT reductions, respectively 15%, and 30%.

  2. 2.

    Vehicle running cost The transition to electric vehicles (EVs) in the context of autonomous travel is expected to lower vehicle running costs (VRC) due to reduced operational expenses associated with EVs. However, the 2022 car cost survey by RACV (2022) reveals that owning an EV may not necessarily result in lower costs when compared to traditional vehicles, because the initial higher purchase price of EVs offsets the advantage of lower running costs. This is likely to happen in the early stage of the AV adoption period, therefore we expect an increase in VRC in AVs. Yet, over time, as AV technology matures and becomes more affordable, AVs are expected to eventually have an affordable price such that they are more economical than a traditional car. We test these possibilities by adjusting VRC, both increasing and decreasing them by 15%.

  3. 3.

    AV penetration From Scenario 2 to 8, a 100% market penetration rate of AVs is assumed among licensed drivers in Clayton, with AVs replacing conventional privately owned vehicles. In addition, AVs do not operate with unlicensed individuals alone or without any passengers. This high level of penetration rate aims to reveal the maximum impact induced by VOTT changes.

  4. 4.

    Activity types Home activities. Fixed activities include work and school activities. Discretionary activities include shop**, leisure, service and walking in natural features, such as parks.

  5. 5.

    Travel modes Four travel modes are included: car driver, car passenger, bike, and walking. Public transportation is not modelled in this study due to several reasons. Firstly, VISTA data reveals that only 8% of trips are attributed to public transportation in Clayton. Secondly, the available public transport modes include trains, trams and buses, making it challenging to obtain detailed travel data for each mode in our high-resolution simulation. Therefore, public transportation is not considered in this study.

  6. 6.

    Accessibility constraints The reduced VOTT is implemented by adjusting the disutility of travel time parameters in choice models. This adjustment allows agents to choose further discretionary activity locations, potentially leading to delays in fixed activities. To address this, the accessibility constraint is applied to ensure that the location choices for discretionary activities do not impede the timely arrival at fixed activities. With different levels of VOTT reductions, agents can adjust the partition within a fixed time budget, as demonstrated in Fig. 5. This diagram describes a one-hour time budget comprising access time, activity duration, and egress time. As VOTT decreases, agents are expected to travel further, resulting in longer access and egress times. To prevent delays in subsequent fixed activities, agents have two approaches to compensate for the extended travel time. The first approach is to reduce the activity duration, while the second approach is to finish the previous activity earlier to maintain the same duration. In reality, the choice between these approaches depends on circumstances. Some activities, such as shop**, can easily vary duration, whereas others, such as medical consultations, are much less flexible. In this case study, we assume that all agents are aware that extended access and egress times necessitate shortening the duration of discretionary activities to avoid delays in fixed activities. The reduction in VOTT is applied proportionally to the reduction in discretionary activity duration. For example, if the VOTT decreases by 30%, the discretionary activity duration also decreases by 30%. As a result, it relaxes the accessibility constraints and enlarges the discretionary location choice set for agents.

  7. 7.

    Network settings In addition to the activity-based travel demand generated by Clayton residents, the study area incorporates traffic originating in other suburbs travelling on the network. This traffic is represented in the form of one-way vehicular trips as the external traffic demand. The distribution of external travel demand throughout the hour of a day is derived from the VISTA data, as depicted in Fig. 6. Together, these two types of traffic demands form 10% of the usual traffic volume in the study area, and therefore, the network capacity is adjusted to 10% in the simulation. This incorporation ensures that the simulation can realistically reflect the real road conditions. “Base scenario validation” section provides the validation of simulation results with real road data.

  8. 8.

    Behaviour constraints The simulation framework models the daily travel behaviour of agents, and certain aspects of behaviour are constrained. Specifically, agents do not undergo relocation of their residential, employment or school locations within the simulation. Moreover, their activity patterns remain consistent across all scenarios, and they do not initiate extra spontaneous activities when utilising AVs.

  9. 9.

    Simulation settings For each scenario, the simulation between mobiTopp and MATSim runs for 3 iterations, wherein MATSim itself undergoes 100 internal iterations to achieve convergence. All simulations are performed on a 32GB RAM, Intel® CoreTM i7-9850H @2.60GHz laptop computer. The running time required for one scenario is approximately 13 h.

Fig. 5
figure 5

Time partition within a one-hour time budget

Full size image
Fig. 6
figure 6

Distribution of external travel demand by hour of a day

Full size image

Results

Base scenario validation

The first part of the results presents the validation process of the base scenario. The activity-based travel demand modelling result indicates that Clayton residents collectively generate approximately 66,000 trips as they engage in their daily activities within the study area. As shown in Fig. 7a, 38% of these trips are for discretionary activities, and 19% are related to work and education. The significant proportion of discretionary trips is attributed to the geographical scope of the study area. Clayton is located 19 km southeast of Melbourne’s central business district (CBD), and our study area excludes the CBD area, which is a prominent employment hub. Figure 7b further demonstrates that more than half of Clayton residents have a distance to work ranging from 10 to 30 km, indicating that the majority of work trips are outside the boundaries of our study area. Trips occurring within the study area by Clayton residents primarily cater to discretionary needs, significantly contributing to the prevalence of this type of trip. Among these trips, leisure activities constitute the majority at 54.2%, followed by shop** trips accounting for 23.4%.

Fig. 7
figure 7

Comparative analysis between the base scenario and 2016 Australian Census data

Full size image

We assess the capability of our simulation framework to represent the aforementioned travel demand. The validation of large-scale activity-based travel demand is always challenging, primarily due to the difficulty in obtaining individual-level data on a comprehensive scale. To address this, we compare aggregated zonal travel time skims with the Uber movement data (Uber Technologies 2023). The Uber movement dataset contains the means and standard deviation of travel time between a pair of origin–destination (OD) zones for each hour of the day, sourced from historical trip records. The delineations of OD zones within the dataset are specific to Uber’s context and are further geographically integrated into our 642 TAZs. We used the 2018 Quarter 3 weekday-only travel times of Melbourne, as the data period predates the impact of COVID-19 and has the fewest public holidays in the year, rendering it particularly representative of traffic on a typical workday. This comparative analysis is based on 10,000 records of Uber OD zonal travel times.

Figure 8a displays a scatter plot comparing the travel times observed from Uber with those estimated by the simulation. The x-axis corresponds to the observed value, and the y-axis represents the estimated value. The solid line indicates the identical value between observation and estimation. The distribution of dots along the solid line indicates their similarity, with only a limited amount of data deviating significantly. Moreover, Fig. 8b further demonstrates the root mean square errors (RMSE) from the observed mean value of travel times between observation and estimation. The RMSE has the following expression:

$$\begin{aligned} \begin{aligned} {RMSE}&= \sqrt{\frac{1}{N}\sum _{1}^{N}\big (t_{pr} - T_{ob}\big )^2},\\ \end{aligned} \end{aligned}$$
(2)

where N is the number of OD pairs, \(t_{pr}\) is the estimated travel time from simulation, and \(T_{ob}\) is the mean travel time extracted from the Uber dataset. To evaluate our RMSE performance, we benchmark against the work by Dey et al. (2020), who adopt a similar method of comparing travel times of each hour of the day using a small Melbourne network. Their RMSE performance between estimated values and Uber data yields a mean of 1.84 and a standard deviation of 0.54. In contrast, our simulation demonstrates a mean RMSE of 3.82 and a standard deviation of 0.69. Given our significantly larger networks and greater traffic volume, we consider our simulation results to be acceptable and successful in representing realistic traffic conditions.

Fig. 8
figure 8

Comparative analysis between the base scenario and Uber movement data

Full size image

Effectiveness of accessibility constraints

The effectiveness of accessibility constraints involves controlling arrival times for fixed activities. This aims to reflect the reality that, despite the presence of AVs, individuals are still constrained by time limitations in their daily routines, which restricts their ability to travel to distant locations as desired. We compare the simulation outcomes of Scenario 6 and 7, where the VOTT is reduced by 30% and the VRC remains unchanged, to evaluate the impact of applying accessibility constraints.

Figure 9a reveals the average trip lengths between the two experimental scenarios. On average, for all discretionary trips, results without constraints show an additional 1.8 km compared to those with accessibility constraints, representing an additional 23% increase in average trip length. Moreover, we observe a 2 km disparity in leisure trips, resulting in approximately a 24% difference in average trip length estimation. Furthermore, we examine agents’ punctuality for fixed activities by comparing arrival time differences in Scenario 6 and 7 with the base scenario. Figure 9b demonstrates that when accessibility constraints are applied, even with a 30% reduction in VOTT, agents can arrive at both types of fixed activities with less than a half-minute delay on average. On the contrary, in the absence of such constraints, simulations allow trips to extend, which causes an average delay of 2.3 and 3.3 min for work and education trips, respectively. Punctuality in fixed activities will still be required even with the availability of AVs. Therefore, the application of accessibility constraints can ensure this punctuality in simulation models, which leads to more robust and reliable results.

Fig. 9
figure 9

Comparative analysis for the effectiveness of accessibility constraints

Full size image

Impact on location choice behaviour

The third part of the results focuses on the impact on discretionary location choices. Table 2 summarises the average trip length of discretionary activity trips by AVs. Our analysis begins with examining all discretionary trips and subsequently delves into specific activity types. The trends observed in the average trip length of all trips align consistently with expectations: as VOTT reduces, average trip length increases; as the VOTT reduction becomes greater, average trip length further extends; and the increase in VRC mitigates the trip length increase from reduced VOTT. These findings correspond with existing literature (Childress et al. 2015; Thakur et al. 2016; Meyer et al. 2017), although the magnitude of increase ranges modestly, between 0.4 and 4%. Our outcomes are more realistic due to the incorporation of accessibility constraints for agents, resulting in more conservative predictions compared to studies that do not account for such constraints. It is reasonable to anticipate a gradual rise in trip length during the adoption period of AVs, rather than anticipating drastic surges in distance travelled.

Turning to specific activity types, the observed trends align with the overall analysis of all trips. A location choice is an outcome of the random utility maximisation process involving factors of locational attractiveness, travel time and travel cost. Diverse marginal utilities for these factors are associated with each activity type, reflecting varying degrees of travellers’ responsiveness to these factors in their location decisions. Moreover, it is important to note that the location choice is made at the zonal level, with the precise location determined randomly after zone selection. We observe that there are minimal changes in trip lengths for service activities. As VOTT decreases, the average trip length for service activities shows negligible growth. This is attributed to the fact that service activities, including medical visits, banking, and other personal business, have the lowest marginal utility of travel time compared to other activity types. Consequently, individuals are less inclined to alter their preferred locations solely due to shorter travel time. This rationale is applicable because individuals tend to visit familiar service providers, such as their preferred medical practitioners, where travel time is less influential in their location choices compared to other activity types. As a result, the average trip length for service activities by AVs remains relatively consistent.

Table 2 Average trip lengths of discretionary activities
Full size table

Impact on traffic conditions

The last part of the results describes the subsequent impact of changes in discretionary location choices on traffic conditions. The integration of MATSim in our simulation framework has facilitated microscopic analysis and visualisation of the traffic dynamics within the study area. Across each scenario, the activity-based and the external travel demand are simulated through 100 iterations to reach convergence in MATSim. We extract the traffic conditions from the final iteration and make comparisons among scenarios. Our analysis uses the average link speed per hour as an indicator, derived from the hourly average travel speed of the entire network. Lower average speeds indicate more network congestion.

Figure 10 illustrates the average travel speeds across all scenarios, yielding two notable observations. First, the average speed demonstrates a declining trend from 5 am, reaching its initial nadir around 8 am. This pattern corresponds to the morning peak hours when most work- and education-related trips embark. Another dip in travel speed occurs around 11 am, primarily attributed to discretionary trips made by residents. Given that the majority of facilities within the study area cater to discretionary needs, this phenomenon aligns with the anticipated surge in errand-related trips around noon. From 3 pm onward, traffic remains consistently busy till 6 pm as individuals commence their journey back home. The traffic congestion begins to alleviate around 7 pm. This depicted pattern mirrors real-world traffic dynamics, thus forming the foundational context for subsequent analysis in regard to the impact of AVs.

The second observation is related to the influence of varying VOTT and VRC on travel speeds, revealing that these changes do not significantly impact the overall travel speed. Specifically, a 15% reduction in VOTT results in minimal changes in average travel speed when compared to the base scenario. Even with a 30% reduction in VOTT and a 15% reduction in VRC, the afternoon peak hours experience a minor decrease in travel speed, albeit with a modest impact. These findings show how different discretionary location choices under accessibility constraints can affect traffic conditions. Given the relatively modest increase in average trip length, we expect that traffic conditions will not significantly worsen. In conclusion, the adoption of AVs can have a negative impact on traffic congestion, aligning with findings from existing literature. However, we expect that traffic conditions will remain similar to the current status during the early adoption period of AVs.

Fig. 10
figure 10

Average travel speeds in various scenarios

Full size image

Implications and discussion

The simulation results conform to our expectations drawn from the existing literature, although they are more conservative. Given the emerging development of AVs and the pressing need for effective transportation and land use planning, we derive two implications from these outcomes.

First, the incorporation of accessibility constraints in our simulation has led to more conservative estimates of trip lengths compared to prior research findings. With the rising use of activity- and agent-based models, the trajectories of individual travel can be reproduced in detail, thereby facilitating the determination of their accessible pathways. Even though AVs offer enhanced travel experiences, it is plausible that individuals still encounter spatio-temporal constraints for work and education-related commitments. Consequently, they might be unable to consistently opt for distant locations with higher attractiveness. By incorporating the accessibility constraints, our simulation framework establishes a theoretically robust foundation for travel behaviour prediction in the era of AVs. It prevents overestimating the gain on trip lengths. Using the accessibility paradigm can help modellers conduct analyses from a holistic perspective, thereby providing more valuable insights for policy deliberations.

Another implication is that while a substantial deterioration in traffic congestion has not been identified due to the extended trip lengths induced by AVs, there can still be a minor decline in average travel speeds if VOTT and VRC are further reduced. Nonetheless, as AV adoption becomes popular and spreads across Melbourne, such a manageable decline in travel speeds is not entirely reassuring. Even though the individual impact of increased trip lengths might seem minor, their cumulative effect could contribute to amplified traffic congestion over time. Furthermore, our simulation restricts AVs from running without passengers, potentially overlooking the presence of additional trips on the roads. As the technology and policies mature, the autonomous relocation function of AVs will be utilised such that on-street parking can be reduced. In that case, empty AV trips could rise and exacerbate traffic congestion (Shafiei et al. 2023). In light of these possibilities, we recommend that policymakers closely monitor the traffic conditions during the early stages of AV adoption and remain prepared to implement policy and infrastructure adjustments.

Conclusion

As our world enters an era of automation, driverless cars are no longer a matter of ‘if’, but ‘when’. The narrative that has been repeated in the literature is that AVs will extend travel distances as individuals are freed from the burden of driving, enabling them to multitask and perceive in-vehicle travel time less negatively. This paper proposed an agent-based simulation framework to assess the impacts of AVs, specifically focusing on variations in VOTT and VRC on discretionary activities. We used Clayton, a populated southeast suburb in Melbourne, Australia as a case study. Our analysis argues that even if individuals travel in AVs, they still encounter spatio-temporal constraints when carrying out daily activities. The outcomes of our simulations reveal that a 30% reduction in the VOTT when using AVs results in only a modest increase in travel distance to discretionary activity locations. This suggests that people are likely to undertake slightly longer trips to fulfil discretionary activities by AVs. Furthermore, our findings indicate that the impact on traffic congestion remains minimal, given the relatively small increase in trip lengths. Our results align with the existing research literature, which suggests that AVs encourage longer trip lengths, thereby validating our simulation work. However, while most studies have concluded that AVs can lead to a significant increase in trip distances, our simulation demonstrates that spatio-temporal constraints can limit distance increments, such that the overall impact of AVs on discretionary location choices is modest. These findings highlight the importance of considering accessibility when predicting the potential impact of AVs, providing policymakers with a more nuanced understanding of the potential effect of AVs on travel behaviour.

Study limitations and future work

Three limitations of the study approach are identified. Firstly, we applied a fixed number of iterations between mobiTopp and MATSim due to the extensive computational running time required by MATSim. Although the exact criteria for convergence between the activity-based model and dynamic traffic assigner remain unclear, existing studies tend to execute a sufficient number of iterations to minimise the possibility of coincidental results. Resolving this issue would require more computational resources to accommodate large-scale simulations. Secondly, the activity patterns used in the case study are based on an existing travel and activity survey, overlooking the potential increased trip rate resulting from the adoption of AVs. As suggested by Truong et al. (2017), AVs can encourage more new trips and increase total daily trips in Melbourne by 4.14%. This indicates that, with AVs available, individuals may change their activity patterns and generate more spontaneous trips. To incorporate this effect, the activity generation process should be redesigned to account for additional trips induced by AVs. Lastly, the lack of empirical data limits information about how many people are willing to travel in AVs. Our assumption in the case study is that all Clayton car owners replace their conventional vehicles with AVs. This assumption may be an oversimplification of reality, as conventional vehicles are unlikely to phase out in the near future. More realistic model assumptions reflecting the penetration rate of AVs are left for future research.

Future work includes incorporating the land use module to facilitate long-term simulations of transport and land use. This study will focus on how AVs reshape land use patterns via accessibility, and how activity patterns of individuals change through a series of long-term location choice decisions.