- 4 The peak spreading phenomenon corresponds to drivers leaving before or after peak travel time to av (...)
1As autonomous vehicles (AV) are becoming more of a reality, with significant technological progress (Pendleton et al., 2017) and numerous experiments (Antonialli, 2019; Stocker and Shaheen, 2019) being led across the world, an increasing number of studies are investigating the economic benefits as well as the costs that can be expected from the development of AV services. These studies have shown that the introduction of such services will have relatively diverse impacts (Narayanan et al., 2020). First, by making the act of driving no longer necessary, users will be able to engage in other activities, such as leisure or work, while sitting in their autonomous car. This is predicted to result in a weakening of the value of travel time savings for private mobility (Correia et al., 2019; Kolarova et al., 2019; Berrada et al., 2020), and subsequently in a reduction in the generalized cost of travel. Regarding public transportation (including taxis and ride-hailing), the absence of drivers is similarly likely to result in lower operating costs as the technology matures (Anderson et al., 2016; Bösch et al., 2018). AV services are also expected to improve accessibility for people with limited motility such as the elderly, children, or adults with no driving license (Meyer et al., 2017). Since the autonomous technology should also result in smoother driving and cooperation between vehicles (e.g., platooning), substantial benefits are also expected in terms of emissions (Bauer et al., 2018), accidents (Clements and Kockelman, 2017), and congestion, as shorter headways between autonomous vehicles could allow road capacity to increase (Simoni et al., 2019). These expected benefits remain controversial, however. AV services might also lead to an increase in traffic due to the lower cost of travel—non-monetary through the cost of time for private transport, monetary through lower fares for public transit—(Fosgerau, 2019; Childress et al., 2015) or due to deadheading (Fagnant and Kockelman, 2014). Lower values of travel time savings could also exacerbate congestion by mitigating the peak spreading phenomenon (van den Berg and Verhoef, 2016).4 Combined with the fact that from a lifecycle perspective, AVs are likely to generate more emissions than conventional electric vehicles due to the additional equipment and data processing they involve, these points make the environmental impacts of AV services highly uncertain (Golbabaei et al., 2020; Wadud et al., 2016). Similarly, there is also strong uncertainty regarding the financial cost of AV services, especially infrastructure costs, which have attracted less attention in the literature.
2Beyond these uncertainties, there is consensus that AV services will result in many changes to transportation supply—new services, lower operating costs—and travel demand—lower value of time, improved accessibility—, together with complex interactions between them (Bahamonde-Birke et al., 2018). In order to better evaluate the (expected) performance of AV services, a growing body of literature therefore relies on mobility simulation models in an attempt to capture these complex factors, ranging from agent-based models—such as MATSim or SimMobility—to direct demand models (Berrada and Leurent, 2017). These studies substantially differ however, be it with regard to the model type used, the way performance is measured, or the services compared within a given study. Many reviews have therefore attempted to synthesize the results of this simulation literature. Berrada and Leurent (2017) provided a short qualitative review of simulation methods and the expected economic impacts (mobility, parking, accidents, environment) of AV services. Jing et al. (2020) carried out a systematic review of the agent-based simulation literature with a corpus of 44 papers, focusing on the simulation platforms used and the critical variables and output of the simulations. Golbabaei et al. (2020) also carried out a systematic review of the literature (with 81 papers) and discussed the expected impacts on urban mobility (fleet size, traffic, and congestion), urban infrastructure and land use (household location, parking spaces, pick-up/drop-off and charging stations), social and travel behavior impacts (trip and mode choice, vehicle ownership) and environmental impacts. Bahamonde-Birke et al. (2018) also discussed what they call the first-order and second-order effects of AV services using a systemic approach. Pernestål and Kristoffersson (2019) reviewed 26 papers and reported their findings on the impacts of AV services, focusing on four specific indicators: the trip (monetary) cost, vehicle kilometers traveled (VKT), fleet size, and waiting time. Other effects have been briefly discussed, such as energy consumption, land-use, and travel behavior. Soteropoulos et al. (2019) carried out a systematic review of 37 modelling studies, with a focus on vehicle kilometers traveled, vehicle hours traveled, modal shares, and land use (parking spaces, including fleet size, location choices). Narayanan et al. (2019) also conducted a comprehensive review of the literature and the reported impacts on traffic and safety, travel behavior, the economy, transport supply, land use, the environment, and governance.
3Their review focused on shared AV services however, in other words, the so-called “robot-taxis”. While several systematic reviews have discussed the expected impacts of AV services, most are qualitative. No meta-analysis has been carried out to date to the best of our knowledge. Furthermore, all the above reviews discuss each impact independently, so the policy implications remain vague due to the wide array of impacts.
4This work aims to better understand our current knowledge—and lack of knowledge—of the expected economic impacts of AV services through a systematic two-step review of the simulation literature. Taking cost-benefit analysis (CBA) as a reference evaluation framework (Layard and Glaister, 1994), we first examine which impacts have been studied in the literature and to what extent by examining the prevalence of 22 indicators directly related to CBA. The cost-benefit analysis seeks to evaluate scenarios (e.g., a new infrastructure, transportation policy or mobility service) by assessing the various impacts, monetizing them, and adding them over time using discount rates in order to determine the value of the scenario for society. While other evaluation methods exist (such as multi-criterion analysis), cost-benefit analysis remains to date the standard evaluation framework for transportation policies across the world (Small and Verhoef, 2007). This first step allowed us to establish a shortlist of key performance indicators for which enough studies were found to carry out a meta-analysis, which was conducted in a second step. While the systematic reviews kept the results attached to the articles in question, the meta-analysis provided overall (i.e., decontextualized) forecasts of the AV impacts. Four key performance indicators (KPIs) were taken into consideration: vehicle kilometers traveled (VKT), travel time, fleet size and total costs. Our meta-analysis thus provides a first quantitative estimate of the expected impact of AV services on travel demand, congestion, and system performance.
5The literature review focuses on road passenger transportation and the (micro-)economic impacts of AV services for society. Applications of AVs to freight (see Flämig, 2016, for a review) are studied in separate papers, with limited (if any) intersection to date with the (passenger) mobility simulation literature. Similarly, autonomous air and rail transportation are not considered in this review due to the specific nature of these modes of transport and the current focus of the simulation literature on road transportation. Our review focuses on (micro-)economic impacts, in other words, all the impacts that may be found in a standard transportation CBA (de Rus et al., 2020). The macroeconomic impacts of autonomous vehicles on economic growth or employment are considered beyond the scope of this paper as they are rarely if ever mentioned in simulation studies and are discussed in other reviews (Clements and Kockelman, 2017; Faisal et al., 2019; Clark et al., 2016). Similarly, other studies offer a broader perspective of autonomous vehicles by considering the state of the art in research as a whole (Gandia et al., 2019), business and management research (Cavazza et al., 2019), and user acceptability (Andersson et al., 2017).
6This paper completes the various systematic reviews on simulation studies in two main ways. First, using cost-benefit analysis as a reference evaluation framework, it quantifies the extent to which listed impacts have been studied in the literature, both separately and jointly. This allows us to show which impacts have been largely investigated and are therefore more likely to be correctly appraised, and which impacts have attracted less attention. By also studying the co-occurrence of impacts -seldom done in former reviews -, we show that comprehensive evaluation of AV services, such as using CBA, remain extremely rare to date, as most modeling studies focus on operational and financial performance with significantly less attention to externalities. Second, this paper provides quantitative rather than qualitative estimates of the expected impact of AVs, depending on the service characteristics for four key performance indicators: VKT, travel time, fleet size and total costs. It thus provides better insights into the effects of AV services on demand, operations and system performance, as well as insights into the influence of service characteristics in this regard.
7This paper investigates the expected economic impacts of autonomous vehicle (AV) services, based on findings in the (passenger) mobility simulation literature. Our methodology relies on two main steps.
8The first step uses a descriptive statistical analysis to determine which impacts are studied in the literature and to what extent. Taking cost-benefit analysis (CBA) as a reference framework for the economic evaluation of mobility services, we establish a list of key performance indicators (KPIs) commonly used in CBA and measure their frequency in a first corpus of AV services modeling studies. In addition to ascertaining the current focus of the relevant literature, this enables us to determine which impacts are studied most frequently and are therefore more likely to be correctly appraised, and which are not.
- 5 The meta-analysis is “a subset of systematic reviews; a method for systematically combining pertine (...)
9Next, we carry out a meta-analysis that focuses on the four most frequently considered KPIs in order to evaluate the expected magnitude of the associated impacts, based on the current state of the art.5 Our first corpus is hence restricted for the sake of the meta-analysis to the subset of relevant studies (i.e., only those featuring at least one of the four KPIs), a subset that we refer to as the second corpus. Through the four KPIs considered (VKT, Travel Time, Fleet Size and Total Costs), the meta-analysis provides a first estimate of the expected impact of AV services on travel demand, congestion, and system performance.
10Figure 1 summarizes the methodology used in the paper, including the data collection process. All of these steps were performed on Excel. Data are available as an online appendix.
Figure 1. Methodology overview
Source: prepared by the authors.
11We now detail the corpus selection process, the nomenclature of AV services used in our study, and the methodologies developed in the descriptive statistical analysis and in the meta-analysis.
12The selection of our two corpuses—the first for the descriptive statistical analysis and the second for the meta-analysis—also comprises two main steps. We began by collecting a preliminary corpus using a standard keyword-based search strategy augmented with a “snowball” search strategy. This corpus was then screened through the successive application of exclusion and inclusion criteria to produce our two final corpuses. As previously mentioned, the second corpus is a strict subset of the first corpus, obtained by considering additional exclusion and inclusion criteria, retaining from the first corpus only studies for which we were able to extract data for the meta-analysis.
13The preliminary corpus was collected using two complementary methods. The primary one was an “All fields” search in the Web of Science database. The keywords used were “(«autonomous vehicles» OR «automated vehicles») AND («Simulation» OR «modelization» OR «modelling» OR «model») AND («passengers» OR «mobility»)”. Only papers published between 1990/01/01 and 2020/10/01 were retained, so one trimester is missing for the year 2020.
14This set of references was extended using a snowball search based on the survey of Berrada and Leurent (2017), which reviews transportation modeling studies on AV services. The snowball search strategy aims to collect a series of papers on a given topic by considering an initial corpus, then expanding it either with the references listed in the corpus (“reverse snowball search”), or with the papers that reference any one of the papers included in the initial corpus (“forward snowball search”) (Francese and Yang, 2021). The starting point can be the result of a search in scientific databases (as in Büchel et al., 2020) or (the solution we chose) an existing review on the topic of interest followed by a search on Google Scholar and Science Direct.
15This twofold search strategy allowed us to obtain a large preliminary corpus of simulation studies about AV services, while limiting possible selection bias inherent to pure snowball search strategies. The Web of Science search offers 529 papers published between 1991 and 2020 (Figure 2). From 1991 and 2013, activity was relatively stable, ranging from one to three papers published a year. From 2014 and 2019, activity grew substantially, from 16 papers published in 2014 to 167 papers in 2019. The number of papers fell in 2020 due to the year being incomplete. Given that one trimester is missing, the number of papers for 2020 should be quite close to that of 2019.
Figure 2. Number of papers per year (preliminary corpus excluding snowball search)
Source: prepared by the authors based on Web of Science query.
16The very low number of studies prior to 2014 may be due to the choice of keywords, as “self-driving” or “driverless” might have resulted in older references. On this point, Gandia et al. (2019) recommend using “Automated” and “Autonomous” when referring to driverless technology.
17We now detail the exclusion and inclusion criteria applied to the preliminary corpus in order to generate the first and second corpuses.
18The exclusion criteria for the first corpus were:
-
Studies dealing primarily with freight;
-
Parking and traffic optimization studies;
-
Studies focusing on the analysis of autonomous rail or air services (including air taxis, technical developments);
-
All papers using simulation in order to provide technical recommendations on computer driving ability. Research driven by the motivation to make autonomous vehicles a technically mature technology, especially in terms of safety, is not included in the scope of our study.
19Conversely, the inclusion criteria for the first corpus were:
-
Mobility simulation studies that consider a scenario with a road-based autonomous vehicle service.
-
The autonomous level considered is SAE level 4 or 5 (see section 2.2 for definitions).
-
Rail services may be included as long as they are only used in the benchmark scenario.
Figure 3. Selection process for the first corpus
Source: prepared by the authors
20The exclusion and inclusion criteria (Figure 3) narrow the number of papers from 529 (preliminary corpus) to 84 articles (first corpus). Simulation and evaluation of AVS is a relatively recent topic in the scientific literature. While the first published paper in the corpus dates from 2014 (Zachariah et al., 2014), almost two thirds of the corpus was published in the last three years (Figure 4), reflecting a growing trend in papers on this topic and evincing the results of the query on the Web of Science.
Figure 4. Publication dates
Source: prepared by the authors.
21The first corpus consists of 42 papers published in peer-reviewed journals, 1 thesis, 40 conference papers, and 1 technical report. Among the conference papers, the International Workshop on Agent-based Mobility, Traffic and Transportation Models, Methodologies and Applications (ABMTRANS conference) is the most represented, with 5 papers (Figure 5). Among the journal papers, the main source is the Transportation Research Record (the Journal of the Transportation Research Board), with 15 papers published. As expected, a fair number of papers originate from the Transportation Research series, with 7 papers in Part A: Policy and Practice, and 7 papers in Part C: Emerging Technologies.
Figure 5. Main publication sources
Source: prepared by the authors. Note: this figure reports all venues with three papers or more.
22The full list of reviews and conferences may be found in the Appendix (Fig. A.1).
23Relative to the first corpus, the following inclusion criteria were added to draw up the second corpus (Figure 6):
-
The study needs to compare two or more mobility services, with at least one involving autonomous vehicles.
-
The paper must evaluate in a quantitative and comparative (across the various scenarios tested in the paper) manner at least one of the four following KPIs:
-
vehicle-kilometers traveled;
-
travel time;
-
fleet size;
-
total costs.
24Conversely, studies from which it was not possible to extract quantitative output of any of the four above-mentioned KPIs were not included in the second corpus. More details on the extraction and treatment of KPIs can be found in subsection 2.4.2.
Figure 6. Selection process for the second corpus selection
Source: prepared by the authors.
25Current and prospective experiments of autonomous vehicles across the world involve a wide array of services, ranging from short-haul on-demand small autonomous vehicles (the so-called robot-taxis, see Stocker and Shasheen, 2018) to autonomous shuttles that operate on conventional stop-based and schedule-based transit lines (see AVENUE for European project). As the specific characteristics of AV services are likely to strongly influence system performance (Nagel et al., 2019), a nomenclature is a useful way to characterize them, in particular allowing us to control for the effect of service characteristics when assessing the impact of AV services in the meta-analysis.
26The following nomenclature was built using the prior work of Antonialli (2019), Földes et al. (2016) and Földes et al. (2018) on smart mobility services. This classification also bears similarities with that of Becker et al. (2020) and Berrada (2019). It is based on the five following features:
-
Vehicle ownership and usage reflects the responsibilities of purchasing, maintaining and sharing the vehicle, and potentially providing a service. We distinguished between two types of ownership: individual ownership, where the vehicle is personal, and third-party ownership, where the vehicle is owned by a public or private operator/organization. Similarly, two types of usage were considered: private usage, where the owner uses the vehicle for his/her own mobility needs, and shared usage, where the owner makes the vehicle available to potential individuals to allow them to reach their destination (Berrada, 2019).
-
The ridesharing feature determines whether the trip may be shared between two passengers or more (Berrada, 2019).
-
Service availability is considered from two perspectives: space (by distinguishing line-based, stop-based, and door-to-door services) and time (on-demand versus scheduled service). This two-dimensional classification is derived from Berrada (2019) and offers similarities with the two operation models described in Antonialli (2019), distinguishing between Regular-Line Transport and Demand-Responsive Transport.
-
Vehicle type mostly refers to the vehicle size, with four increasing levels of capacity. The “car category” is used for vehicles with 1 to 5 available seats, shuttles for 6 to 18 seats, and buses for more than 19 seats. The rail vehicle type includes tramways, metros, and trains (inspired by Stocker and Shaheen, 2017).
-
Automation level describes the vehicles’ automation features based on the SAE classification (SAE J3016:201806 “International Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles”). Three main levels are explored in the literature: conventional level refers to the level 0 of automation, semi-autonomous level to levels 3 and 4, and autonomous level to level 5.
Table 1. Strategic features of a mobility service
|
Service nomenclature
|
Features
|
Source
|
|
Vehicle ownership and usage
|
Individual ownership and private usage Individual ownership and shared usage Third-party ownership and shared usage
|
Berrada (2019)
|
|
Ridesharing
|
Yes No
|
Berrada (2019)
|
|
Service availability
|
Space
|
|
|
Door-to-Door Stop-Based Line-based
|
Hardt and Bogenberger (2016)
Berrada (2019)
Antonialli (2019)
|
|
Time
|
|
|
Scheduled On-demand
|
Antonialli (2019)
|
|
Vehicle type
|
Car Shuttle Bus Rail
|
Stocker and Shaheen (2017)
|
|
Automation level
|
Conventional (Levels 0,1,2) Semi-autonomous (Levels 3&4) Autonomous (Level 5)
|
SAE classification
|
Source: prepared by the authors based on above sources.
27The on-demand door-to-door system collects passengers from their location and takes them to their final destination. The vehicle can either be shared (ridesharing) or used privately (private ownership or solo car sharing).
28The on-demand Stop-Based system is a hybrid between conventional public transit and on-demand door-to-door services. Boarding/alighting is only permitted at stations. Again, the vehicle can be either shared or used privately.
29The descriptive statistical analysis aims to determine which impacts are considered in the simulation literature, and to what extent, at the same time allowing us to highlight the current focus of the literature and the gaps to be filled.
30In order to list which impacts are considered in simulation studies, we took the cost-benefit analysis as the reference evaluation framework, since it is currently standard practice in carrying out an economic evaluation of transportation investments, services, or policies (Boardman, 2006; de Rus et al., 2020). We then considered a list of common CBA indicators (de Rus et al., 2020; Quinet, 2013; Small and Verhoef, 2007; Victoria Transport Institute, 2009) and classified them into five main categories, related to:
-
Service performance: includes indicators that measure service performance from a demand perspective, which are then used to compute consumer surplus,
-
Operations: includes indicators that measure the operators’ economic performance, which is then used to compute the operator surplus,
-
Externalities: includes the main externalities captured in standard CBA: i.e., energy, greenhouse gas emissions, local pollutants, noise, safety, congestion,
-
Socioeconomic: used to determine whether the study evaluates the results through the prism of some socioeconomic characteristics such as age or income level,
-
Cost Benefit Analysis (CBA): refers to the standard final output of a CBA, such as the net present value.
31Table 2 lists the indicators (twenty-two) identified within each category. The aim of this classification is to highlight the focus of the studies, while evaluating the capacity of the different simulation models used in the papers included. Apart from the “Socioeconomic” category, these indicators were selected for their use in CBA. The “Socioeconomic” section aims to complete the picture by adding social indicators where a few other studies, such as Tian et al. (2018), had a more technological-oriented indicator set (such as pre-collision systems or machine learning approach-based emergency brakes).
32The selected indicators are usually simulation outputs. In some specific cases, they may also be considered as operational constraints to ensure a certain quality of service: e.g., fleet size (Berrada, 2019; Vosooghi, 2019) or the required Level of Service (Navidi et al., 2017).
Table 2. Indicators
|
Category
|
Indicator
|
Definition
|
Source
|
|
Service performance
|
Waiting time
|
|
Quinet (2013), De Rus et al. (2020)
|
|
Patronage
|
|
De Rus et al. (2020)
|
|
Travel Time
|
Total travel time is composed of (a) access/egress time (only for line-based and stop-based services), (b) waiting time, and (c) in-vehicle travel time
|
De Rus et al. (2020)
|
|
Vehicle Kilometers travelled (VKT)
|
Refers to the total distance travelled by empty and loaded vehicles. Also measured by vehicle miles travelled (VMT) in the Anglo-Saxon literature
|
Quinet (2013), Small (2007)
|
|
Operations
|
Occupancy rate
|
Corresponds to the average number of passengers per vehicle
|
Huang et al. (2020) *
|
|
Fleet size
|
Refers to the total number of vehicles deployed for the service production
|
Bösch et al. (2016)
|
|
Total Costs
|
Production costs, including at least operating costs. Fixed costs are also included in some studies
|
Bösch et al. (2018)
|
|
Profits
|
Corresponds to the net profit for the operator, as the difference between costs and revenue
|
Fagnant and Kockelman (2016)
|
|
Fare
|
Corresponds to the price of usage
|
Tirachini and Antoniou (2020)
|
|
Externalities
|
Local pollution
|
Corresponds to pollutant emissions that contribute to poor air quality, including particulate matter (PM), nitrogen oxides (NOx), and volatile organic compounds (VOCs)
|
De Rus et al. (2020)
|
|
GHG emissions
|
Mainly include CO2 emissions due to operations. Refers to climate change
|
De Rus et al. (2020)
|
|
Energy
|
Corresponds to energy consumption required for operations according to the motorization type of vehicle (thermal, electric, etc.)
|
Bauer et al. (2018)
|
|
Noise
|
Refers to consideration of the noise nuisance
|
De Rus et al. (2020)
|
|
Safety
|
Corresponds to the level of accidents with injuries and deaths
|
De Rus et al. (2020)
|
|
Congestion
|
Corresponds to an estimation of the resources wasted in an overcrowded environment
|
De Rus et al. (2020)
|
|
Socioeconomic
profile
|
Age
|
|
Urbina and Sohaee (2020)
|
|
Gender
|
|
Hulse et al. (2018)
|
|
Socioeconomic status
|
Variables describing the socioeconomic status of the household/individual, such as income, job category…
|
Müller et al. (2020)
|
|
People with reduced mobility
|
|
|
|
Cost
Benefit
Analysis
|
Net Present Value (NPV)
|
Corresponds to the difference between the present value of inflows and outflows over a certain period of time
|
Quinet (2013)
|
|
Internal Rate of Return (IRR)
|
The annual growth rate an investment is expected to generate
|
Quinet (2013)
|
|
Benefit Cost Ratio (BCR)
|
The ratio between costs and benefits, expressed in monetary or qualitative terms
|
Quinet (2013)
|
* The “occupancy rate” refers to the “average vehicle occupancy” from Huang et al. (2020).
Source: prepared by the authors based on above sources.
33We investigated the use of indicators in the literature with a statistical descriptive analysis examining 1) the occurrence of each indicator, 2) the occurrences of at least one indicator per category, and 3) the mean numbers of indicators per category in each paper. The analysis is followed by a qualitative discussion regarding the use of indicators in the papers reviewed.
34In addition to the use of indicators, our analysis also investigated the types of simulation models used, providing some insights into the capacity of the various model types to simulate specific AV services or to generate distinctive outputs.
35The following model types were considered, using the standard classification of transportation models (Ortúzar and Willumsen, 2011; Soteropoulos et al., 2019):
-
Agent-based models are models where “a system is modeled as a collection of autonomous decision-making entities called agents. Each agent individually assesses its situation and makes decisions on the basis of a set of rules.” Bonabeau (2002)
-
Four-step models are traditional mobility simulation models. They offer an aggregated view of the demand and supply of mobility (McNally, 2007).
-
“Direct demand models can be of two types: purely direct, which use a single estimated equation to relate travel demand directly to mode, journey and person attributes; and a quasi-direct approach which employs a form of separation between mode split and total (O–D) travel demand. Direct demand models are closely related to general econometric models of demand and have long been inspired by research in that area.” (Ortúzar and Willumsen, 2011, chapter 12; Talvitie, 1973)
-
Traffic models represent road traffic flows based on the vehicles’ capacity to interact with each other and the infrastructure.
-
Land Use/Transport Interaction “illustrates the spatial organization of the network of socio-economic activities and describes the physical separation between them. The transportation system connects the various activities/land uses” Gavanas et al. (2016)
-
Fleet control models are supply-focused models providing rules to assign vehicles to their goals.
-
Mode choice models assign travel demand to specific modes according to their socioeconomic parameters (Ortúzar and Willumsen, 2011, chapter 6).
36The meta-analysis completed the descriptive statistical analysis by providing a quantitative estimate of the (expected) impacts of AV services, focusing on the KPIs for which enough studies were reported in the statistical analysis.
37According to the descriptive statistical analysis, of the twenty-two indicators scrutinized, four KPIs stand out in terms of coverage and quality of treatment.
38First, the level of demand was assessed with the Vehicle Kilometers/Miles Traveled (VKT or VMT). This is a paramount KPI for mobility services as it not only relates to travel demand, but also to traffic flows, operating costs and revenues, and virtually all externalities. It is therefore a key driver of economic profitability in a CBA (Small and Verhoef, 2007).
39Travel time is a key performance indicator used to measure service quality. Travel time (TT) refers to the total travel time composed of (i) access/egress time (only for line-based or stop-based services), (ii) waiting time, and (iii) in-vehicle travel time. Access/egress time and waiting time may be equal to zero for some services, such as the private car. Travel time is strongly related to consumer surplus and thus to economic profitability (Small, 2012). Conversely, as the service evaluated is autonomous, it has a much smaller impact on operating costs since no drivers are involved (Tirachini and Antoniou, 2020). Similarly, as most AVs are electric, the impact of travel time (or more specifically of speed) on the environmental externalities is greatly reduced.
40Third, fleet size is a widespread key performance indicator in the literature, especially for on-demand services, as many papers study the link between fleet size, dispatch strategies and operational performance (e.g., Hörl et al., 2021). While it is strongly related to capital costs, it also exerts influence on traffic, congestion and/or parking needs, and environmental externalities (from a lifecycle perspective).
- 6 The total cost of ownership approach is actually highly congruent with the cost-benefit analysis in (...)
41Finally, the total cost of a mobility service measures the total production cost of the service, including at least the operating costs and capital costs if available. Unlike Tirachini and Antoniou (2020), we did not consider user costs, as these were already captured by the travel time indicator. In the simulation literature, costs are generally averages per year that are typically derived from more precise total cost of ownership approaches that measure the cost of a vehicle over its entire lifecycle, including purchase, operations, and end of life (e.g., Ongel et al., 2019).6 For AV public transit services (including taxi and ride-hailing), the absence of drivers could lead to a decrease in operating costs (Bösch et al., 2018). Conversely, if the AV service attracts more demand than the former conventional service, Total Costs could rise as a result. The total costs indicator thus reflects both operational and marketing performance. The choice of a total costs indicator relative to unit cost indicators (cost/km, cost/seat, or cost/passenger) relates to 1) the interest of the study of overall service performance and 2) the constitution of the cost indicators. Unit costs are frequently model inputs, often based on Bösch et al. (2018), whereas total costs reflect output from the simulation or result from an economic evaluation based on the simulation output.
42The performance of the autonomous vehicle service was measured relative to another service. For each of the four KPIs (VKT, Travel Time, Fleet Size, Total Costs), we computed the mean relative variation between the reference service and the service compared. Consider for instance an AV service providing rideshared door-to-door trips, and that the VKT for this service is found to be equal to 115% that of the reference service (conventional private cars, for instance). This means that when these two services are compared in the meta-analysis, the AV taxi service will result in a +15% increase in VKT relative to conventional cars. While the usual benchmark against AV services is conventional private vehicles, some studies test AV services against conventional public transit or other AV services (e.g., by comparing AV services that offer private trips versus rideshare trips).
43Studies may consider several scenarios regarding the service characteristics or the economic environment (e.g., market penetration, adoption levels). When scenarios consider different shares of Avs within the vehicle fleet, we decided to keep only the scenario with the highest penetration ratio. For instance, Llorca et al. (2017) considered two market penetration scenarios, with 20% and 40% share of Avs within the vehicle fleet, respectively. In this case, only the 40% scenario was retained. If more than one scenario with a high penetration rate is considered in a study, the KPI is averaged across the scenarios. In some cases, developed below, the performance variation needed to be estimated to obtain a proxy of the service implementation impact.
44To illustrate this methodology, we put forward our main assumptions and the corresponding examples below by considering VKT as the indicator being evaluated. The methodology is similar for Travel Time, Fleet Size and Total Costs. If the scenarios involve varying fare levels, substitution rates of AV/per trip, or modal shares, then the indicator was averaged across scenarios.
45In cases where the KPI evolution was combined with other modes, a ratio was used to estimate the KPI evolution. In Oh et al. (2020), VKT are estimated for two adoption level assumptions: a) a High adoption scenario and b) a Moderate adoption scenario. Again, only scenario a) was used. Moreover, the VKT for AMOD (Autonomous Mobility On Demand), which encompasses AV (autonomous taxi) and SAV (shared autonomous taxi) services, was computed as a single synthetic mode (Fig. 7). Thus, to differentiate the performance of the two modes, a ratio from their respective modal shares was used to assign a proxy of the VKT of each mode.
Figure 7. Results of simulation for vehicle kilometers travelled
Source: Figure 13 from Oh et al. (2020)
46The average rise in AV and SAV VKT were extracted from the average of AMOD VKT (4), in other words, AV + SAV VKT, with the average of the 3 price scenarios of the high adoption scenario, in other words: (+42% + 32 % + 25%)/3 = + 33%, with the average ratio of modal share for AV (6.27%) and SAV (8.83%) on total AMOD modal share (15.10%) from Fig. 8.
Figure 8. Simulation results for modal shifts
Source : Oh et al. (2020), Figure 12
47The average share of the rise of VKT from the AV and SAV is respectively: 33% x 6.27 % / 15.1% = +14% and 33% x 8.83%/15.1% = +19%.
48The two categories of indicators investigated in all 84 papers of our first corpus (apart from one paper on operations) cover service performance and operations (Figure 9). They also feature the highest average number of indicators per paper. For service performance, the mean number of indicators per paper is 3.37 (from 4 possible indicators), and only 12 papers (14%) use less than 3 of these indicators. Surprisingly, the VKT indicator is the least represented, yet still has over 76% of occurrence (Figure A.2A in Appendix). The VKT indicator may be straightforwardly computed by multiplying total ridership by the mean travel distance, which might explain why it is not always reported. Regarding operations, the average number of indicators per paper was 2.64 (from 5 possible indicators), and 43 papers (51%) use fewer than 3 of these indicators. Fare and profit are the least represented within this category, while fleet size, total cost and utilization rate are present in respectively 83%, 73% and 55% of the corpus (Figure A.2B in Appendix). In fact, a large number of papers deal with optimal fleet size and dispatching strategies (e.g., Fagnant and Kockelman, 2016; Loeb and Kockelman, 2019; Vosooghi et al., 2019). The latter three indicators—fleet size, utilization rate, and total costs—tend to form the crux of the analysis, while fares and profits may be disregarded in that they are more related to demand.
49Externality indicators are considered at least once in fewer than two-thirds of the papers reviewed (68% of occurrence). Usage is more heterogeneous than for the two previous types of indicators (Figure A.2C in Appendix): the prevalent KPI within the category is congestion (61% of occurrence), followed by energy (33%) and climate change (26% of occurrence). Local pollution (21% of occurrence) always appears together with a climate change indicator, reflecting the fact that no paper in our corpus focuses on air quality, while climate change is given slightly more attention. Noise and safety KPIs have significantly lower occurrence rates (1% and 6% respectively), and again always appear in combination with climate change KPIs (e.g. Simoni et al., 2019).
50Socioeconomic profile indicators appear at least once in 21% of the corpus, with 0.31 indicators per paper on average. Moreover, only four papers include more than one socioeconomic category indicator (e.g. Berrada, 2019).
51Finally, just two papers in the whole corpus (2% of the corpus) include CBA indicators, with a single indicator each time (Figure A.2D in Appendix): either the BCR (Gelauff et al., 2019) or the IRR (Fagnant and Kockelman, 2016).
Figure 9. Occurrence of indicators per category
Source: prepared by the authors.
52The analysis shows a very strong prevalence of service performance and operational indicators, reflecting the fact that most papers focus on performance from the perspective of the user or the operator, but seldom that of society as a whole. Since most AVS tested in the papers are on demand, fleet control and optimal dispatching strategies attract considerable interest (again, to cite just a few, Ben-Dor et al., 2019; Fagnant and Kockelman, 2016; Farhan and Chen, 2018; Loeb and Kockelman, 2019; Vosooghi et al., 2019). As result, the most frequently investigated KPIs are supply oriented, including, on the operator side, the VKT, fleet size, utilization rate, and costs, and on the user side, waiting time and travel time. Conversely, KPIs that are demand-oriented such as fares or more elaborate KPIs such as profit are covered less in the corpus.
53In addition to being underrepresented compared to service performance and operational KPIs, the indicators relating to externalities tend to be those where the analysis is the least thorough. In some papers, pollutant emissions (GHG and local pollutants) and congestion are mentioned but not analyzed (Simoni et al., 2019; Zachariah et al., 2014; Heilig et al., 2017, Jäger et al., 2018). In others (Navidi et al., 2018; Wang et al., 2018; Childress et al., 2015), pollutant emissions and congestion are repeatedly proxied by the distance traveled, which indicates the general trend but not the intensity of the trend. Furthermore, while VKT is a key driver of pollutant emissions, vehicle type (especially if the AV service is electric), average speed and congestion also have a strong influence (Grote et al., 2016), which is not captured when VKT is used as a proxy for emissions.
54Socioeconomic category indicators (age, gender, income class, people with reduced mobility) are often mentioned but seldom analyzed. For instance, people with reduced mobility are mentioned once (Sieber et al., 2018), but they do not fulfill any role in the simulation process. In Meyer et al. (2017), age acts as an important parameter of the demand simulation with AV taxis since people with limited access to mobility increase the overall demand by 16%. Similarly, Puylaert et al. (2018) do not distinguish travel behavior between age or social categories, but include these parameters to determine the type of car owned. Truong et al. (2017) proceed in a similar way. In these papers, age only serves as a segmentation variable in the demand model. Only one paper carries an in-depth analysis by investigating the impact of AV service on the mobility of the age brackets with mobility issues (Kamel et al., 2018).
55Cost-benefit analysis indicators are extremely infrequent: only two papers (2% of the corpus) feature them, with a single indicator in each case. Moreover, in Fagnant and Kockelman (2016), the IRR is used as a financial indicator, meaning that only the study of Gelauff et al. (2019) actually engages in a welfare analysis of AV services. Moreover, the latter study focuses on consumer surplus, and does not consider either the operator surplus or externalities such as congestion, safety and pollutant emissions. This virtual absence of cost-benefit analysis indicators is not surprising if we consider that CBA is a step further from environmental indicators which are already poorly represented within the corpus (with only 21% of occurrence). This confirms that the focus of the AV simulation literature is currently strongly oriented toward the operational design of AV services (including fleet size, dispatch and pooling strategies), rather than their strategic design, which would involve a welfare analysis (in most cases involving the computation of CBA indicators).
56Agent-based models (ABMs) represent the large majority (75% of the corpus) of the models used for AV simulation (Table 3). Within this category, the MATSim open-source framework is used in more than half the agent-based model papers (41%). While in most cases, including all MATSim instances, travel demand is determined endogenously, some ABMs tend to treat travel demand as exogenous, and focus on fleet control. The agent-based modeling paradigm is then used to consider user-vehicle interactions regarding waiting times and/or pooling decisions in the matching process between users and vehicles.
57Four-step models represent the second largest category, yet account for only 10% of the second corpus. Activity-based (i.e., non agent-based ones) models are an intermediate form between agent-based and four-step models: while their representation of demand is identical to that of activity-based ABMs, the transportation supply is represented in a simpler and more aggregate manner than in four-step models. This aggregate representation of supply precludes representing vehicle dispatching strategies, and thus finely analyzes the operational performance of the AV service. When four-step or activity-based models are used, the performance of the AV service is often evaluated through the lens of modal shares (Levin and Boyles, 2015) or trip characteristics, such as the mean trip length or duration (Childress et al., 2015; Zhao and Kockelman, 2018). Berrada (2019) is an exception as in this case a four-step model running in VISUM is coupled with VIPSIM, an agent-based fleet control-oriented model, resulting in a range of indicators closer to ABMs than to four-step and activity-based models.
58Direct Demand Models are the third largest category of models, representing 6% of the papers. These models are used to generate travel demand based on supply and demand characteristics, but with no or very limited representation of spatial interactions (Anderson et al., 2006). They are often used when considering aggregate trips at the level of a country, a region, or a specific origin-destination (Ortúzar and Willumsen, 2011).
Table 3. Models used in the first corpus
|
Models
|
Use
|
|
Agent-based
|
75%
|
|
Four Steps
|
10%
|
|
Direct Demand model
|
6%
|
|
Traffic model
|
2%
|
|
LUTI
|
2%
|
|
Fleet control model
|
2%
|
|
Mode choice model
|
2%
|
Source: prepared by the authors
59Other models used in the corpus were only exploited twice, but not enough papers were gathered to provide conclusive evidence. This category included the use of Discrete Choice Models. Like direct demand models, these models are often preferred to four-step models—which add the generation, distribution and assignment steps to the mode choice—when spatial interactions and network effects (as when users switch from one service to another, since a change in the transportation supply or in travel conditions leads to a new supply-demand equilibrium) are not the focus of the paper. For instance, Truong et al. (2017) provide a rough estimate of the impacts of AVS in Victoria, Australia, with no spatialization of the results. Sun et al. (2020) estimate a mixed-logit mode choice model to investigate user preferences and assess whether cost savings or travel time savings are more important for users when comparing AVS with Conventional Public Transit services, meaning that, once again, spatial interactions are not a central issue. Other models were also identified in our corpus that target very specific issues that are not part of our study. For instance, LUTI (Land Use and Transport Integrated) models highlight the interaction between land use (mostly job locations and residential areas) and transportation. Fleet control models are a type of model that focus on the supply side of transportation. When simulating on-demand services, ABMs often use a fleet control module, such as the DRT module for MATSim. Lastly the traffic model provides an analysis of interactions between infrastructure and vehicles through the infrastructure characteristics and the vehicle capacities. There is less focus on the service level than on the infrastructure level, which is mainly why this type of model does not appear much in our corpus.
60There is thus a strong focus on agent-based models in the literature. These models emphasize analysis at tactical level, with relatively close attention to fleet optimization (compared to the four-step models). They aim to optimize operational efficiency by maximizing the utilization rate (Fagnant and Kockelman, 2016; Vosooghi and al., 2019; Llorca et al., 2017) and/or the level of service for a given fleet (Wang et al., 2018; Lu et al., 2018).
61Agent-based models are more suitable to describe dynamic availability of mobility services and interactions between agents compared to four-step models. As mentioned in 2.2, the corpus is primarily composed of on-demand services: of the 63 studies made with agent-based models, 61 consider on-demand services. Thus, it is no surprise to see such widespread use of these models in the corpus.
62In addition, studies that use agent-based models do not take different classes of demand into consideration (depending on their socioeconomic profile, for instance), which represents a future research path, especially since AVS show promise in improving accessibility of age brackets with mobility issues.
63For the second corpus, the papers selected use a least one of the four KPIs (VKT, travel time, fleet size and/or total costs) in a comparative form. The number of documents dropped from 84 documents in the first corpus to 48 documents in the second. The total number of services compared is 80, however, since several papers evaluate more than one AVS.
64Based on the service nomenclature defined in 2.2, we generated combinations of services that were investigated in our second corpus. Ten combinations of services were finally identified and named, as presented below in Table 4.
Table 4. List of services considered in the meta-analysis
Source: prepared by the authors.
- 7 Note that the methodology adopted for data extraction, which averages indicators in the case of mul (...)
65The performance comparisons between autonomous and conventional services show strong differences depending on the service of reference. When compared to private cars, autonomous services offer overall performance variations of +/- 50% on the four indicators (Figure 10). The variations are much larger (up to +700%) when the service of reference considered is Conventional Public Transport (Figure 11).7
66The number of papers covering comparisons between two (or more) services also indicates the interest of the academic community in such comparisons. In the corpus collected, the three most frequently compared service pairs are:
-
the replacement of private cars by Private AV,
-
the replacement of private cars by autonomous vehicles (AV), corresponding to autonomous taxis (shared vehicles but no ridesharing),
-
- 8 There is no clear consensus in the literature on the definition of “Shared Autonomous Vehicles”, as (...)
the replacement of private cars by shared autonomous vehicles (SAV).8
Figure 10. AV On-demand service performance against conventional counterparts
Acronyms: VKT = Vehicle Kilometers Traveled, TT = Travel Time, FS = Fleet Size, TC = Total Costs
Example of interpretation: here the first comparison on the left is the performance of Private AV vs. Private Car. The four indicators (VKT, Travel Time, Fleet Size and Total Costs) are covered by the literature, and the replacement of private cars by private Avs should result, on average, in a rise of VKT and Travel Time (by respectively +17% and +7%), but also in a reduction of Fleet Size and Total Costs (respectively by 10% and 17%).
Source: prepared by the authors.
67These three categories represent the majority of comparisons. The next major comparison is between conventional public transit and AV/SAV (~15% of comparisons).
68According to Stocker and Shaheen (2019), AV and SAV services are the main autonomous vehicle business models projected by the major manufacturers (Ford, Tesla, Daimler) and tech developers (Google or Uber). It is not surprising that academic attention also focuses on these service types. On the other hand, shuttle-based services are explored to a greater extent by public transport operators via several experiments worldwide (see SAM project).
Figure 11. AV service performance against Conventional Public Transport
Source: prepared by the authors.
69Focusing first on the comparison between AV-based services and private cars, it clearly appears that the three most studied AVS (Private AV, AV and SAV) are three steps from the same ladder, trading the VKT and travel time performance of private cars against their fleet size performance at different intensities (Table 5).
Table 5. Performance comparison of conventional and autonomous services
|
Reference
|
New Service
|
VKT
|
Travel Time
|
Fleet Size
|
Total Costs
|
|
Private car
|
Private AV
|
+17%
|
+7%
|
-10%
|
-17%
|
|
Private car
|
AV
|
+23%
|
+17%
|
-17%
|
N/A
|
|
Private car
|
SAV
|
+6%
|
+20%
|
-55%
|
N/A
|
|
Public transit
|
AV
|
+464%
|
-52%
|
+727%
|
-26%
|
|
Public transit
|
SAV
|
+361%
|
-32%
|
+377%
|
-18%
|
|
AV
|
SAV
|
-16%
|
-8%
|
No variations
|
-22%
|
*Regarding comparisons with public transit, SAV is found to be more costly than AV (-18% versus -26%), based on 3 and 4 occurrences respectively. On the other hand, the direct comparison of AV and SAV leads to the opposite result, in other words, SAV is less costly (-22%) but based on only one occurrence. These results should thus be considered with care.
Source: prepared by the authors.
70Private AV services offer the best performance regarding Travel Time, which increases by only +7% compared to Private Car, as opposed to +17% for AV and +20% for SAV. This is mostly due to the fact that private AV do not involve waiting time as is the case for AV or SAV (Fagnant et al., 2016). If VKT and Travel Time are greater for private AVs than for Private cars (+17% and +7% respectively), it is probably because the marginal generalized cost of the former is lower than that of the latter. The expected decrease in the value of travel time savings from not having to drive (Kolarova and al., 2019; Fosgerau, 2019; Singleton, 2019; Szimba and Hartmann, 2020; Gao et al., 2019) should result in both more frequent and longer trips.
71The marginal operating cost could also be lower for autonomous vehicles operated by a third party than for conventional ones. This should have a considerable impact on mobility services in which drivers’ wages are an important component (Bösch et al., 2018; Bauer et al., 2018; Loeb and Kockelman, 2019). However, these analyses rely on forecasts not confirmed as yet by empirical data. Similarly, the cost analysis of Leich and Bischoff (2018) is based on assumptions that cannot be validated based on real data, which would have a significant impact on their findings.
72The effect on VKT is relatively similar for AV (+23%) and SAV (+6%) as for Private AV (+17%), but the fleet size is smaller due to ridesharing. The effect on fleet size is even greater for SAV (-55%) than for AV (-17%) since ridesharing also makes vehicles not fully loaded available to passengers (Farhan and Chen, 2018). Similarly, VKT are lower for SAV than for AV (-16%) since fewer vehicles are assigned through a centralized dispatcher, while maximizing their loading. In addition, SAV would attract at least the same number of passengers since they usually charge lower fares than AV (Simoni et al., 2019; Vosooghi et al., 2019), reducing their total generalized cost, even for less comfort and additional detours (Golbabaei et al., 2020). Ridesharing can also help to further reduce waiting times, especially during peak times, by increasing vehicle availability (Hörl, 2017). However, given that travel times include waiting times as well, the literature notes that the increase in travel time is greater for SAV (+20%) than for AV (+17%), indicating that the extra Travel Time resulting from detours exceeds the reduction in waiting times (Vosooghi et al., 2019; Farhan and Chen, 2018).
73Regarding vehicles’ capacity, the optimal capacity providing shared door-to-door trips is found to be between two and four seats per vehicle (Leich and Bischoff, 2018; Berrada, 2019; Zachariah et al., 2014; Gurumurthy et al., 2019; Farhan and Chen, 2018; Vosooghi et al., 2019). In fact, in general, the average occupancy is found to be about two persons per vehicle. These numbers should be taken with caution. The effect of empty kilometers traveled was not assessed in this meta-analysis and their role in the occupancy rate might be important. The average occupancy per vehicle also decreases with fleet size (Winter et al., 2018). That being said, Wang et al. (2018) published a paper on the ridesharing potential of Singapore, based on real taxi booking data, where 40% of the trips were shared by six passengers or more in taxis, showing the potential of this type of service in densely populated urban areas. Winter et al. (2018) and Navidi et al. (2017) also presented evidence of the ability of AV and SAV to take advantage of economies of scale, even if the leverage seems weaker than the conventional public transit leverage. This shows that benefits from a reduced Fleet Size are more an outcome of sharing vehicles (sequentially or simultaneously) than the effect of automation. Zhu (2019) exposed that the extrinsic monetary incentive did not provide leverage to support ridesharing policies. The automation innovation might help to promote societal changes.
74We now turn our attention to the comparison between conventional public transit and autonomous services. Shuttle-based services provide a vehicle capacity of eight to fifteen seats, allowing more passengers to board than car-based services, while offering greater flexibility than conventional public transit. In this configuration, a shuttle service would save passengers time (Sieber et al., 2020; Bischoff et al., 2018; Viergutz and Schmidt, 2019), but would also offer both the service provider and passengers savings at the cost of a larger fleet (in number of vehicles, though not of the same size). This topic has attracted less attention in the academic community and even fewer studies that would correspond to the methodology established for the second corpus.
75From our results (Figure 11), stop-based routing seems to be the best autonomous alternative to conventional public transit to limit externalities (here proxied by fleet size and VKT). In addition, a stop-based SAV service is also more likely to benefit from economies of scale than a door-to-door shuttle (fewer detours, thus shorter travel times and waiting times, and less congestion).
76The comparison of AVS with conventional public transit suggests that AVS could be interesting in peri-urban or rural areas where conventional public transit might struggle to benefit from economies of scale, or as a feeder (first mile and last mile) service. The expected decrease in operating costs from the drivers’ salaries could allow smaller and more flexible vehicles to operate, reducing both passenger waiting time and the overall system costs (Berrada and Poulhes, 2021; Schlüter et al., 2021). Another alternative is to operate buses with a higher level of service but reduced capacity (Bösch et al., 2018; Winter et al., 2018). However, these reductions come at the price of a larger fleet size and more VKT (Sieber et al., 2019; Bischoff et al., 2018; Leich and Bischoff, 2018; Bösch et al., 2018b; Merlin, 2017; International Transport Forum, 2015; Imhof et al., 2020) which might be less of a problem in rural areas, where the space dedicated to mobility has a lower opportunity cost, and the externalities generated by transport are less of an issue than in urban areas. In addition, most of the publications in our corpus focus on urban and peri-urban areas, with only two articles specifically dealing with rural areas (Viergutz and Schmidt, 2019; Sieber et al., 2020).
77However, the study by Leich and Bischoff (2018) warns of the dangers of competition between conventional public transit and AVS, since AVS would take over the public transit passengers and reduce its profitability. This could result in a reduction in the level of service for public transit, which would exacerbate the modal shift toward AV services. The study of ride hailing firms by Carballa Smichowski (2018) shows that the ride hailing companies might be prone to use predator price, which would strengthen the threat if these companies were the ones developing the autonomous vehicles. The authors therefore recommend conventional public transit operators switching to autonomous vehicles. Hatzenbühler et al. (2020) compared the performance of conventional and autonomous buses and found that autonomous bus services were less expensive to operate and provided better travel times to boot.
78The idea of public transit lines being substituted by autonomous flexible vehicles is discussed not only in the academic field, but also in private industry. Rau et al. (2020) investigated the effect of autonomous shuttle pods driving through Singapore in dedicated lanes. These pods could be linked together to create little trains to absorb demand during peak hours and then divided into multiple vehicles to provide attractive Level of Service, even during off-peak times. This idea is similar to the Loop Aix-Marseille project and some Hyperloop projects (see Loop Aix Marseille and Urbanloop references).
79The regulator is the actor interested in the analysis of the externalities. As a complement to simulation models, tools have been developed (cost-benefit analysis and multicriteria analysis) to assess these issues. However, very few studies offer an analytical framework that allows the economic appraisal of AV to be carried out in situ.
80In order to make the best strategic choice (i.e., to define the “best” mobility service with respect to the given mobility and sustainability objectives), we cannot rely on an eclectic set of indicators. The regulator needs a comparative base between projects, with the potential to rank them. The average methodology could overshadow concerns about the heterogeneity of demand during the day (peak and off-peak time), the type of territory (urban, peri-urban and rural), or the field of application of the mobility service (First and Last Mile service, for example).
81The sample of articles provides interesting opposing trends between Private Car and Private AV, AV and SAV services, but the comparison of other service pairs is less robust. If autonomous taxis (AV and/or SAV) were to replace conventional private cars, they should increase VKT between +23% and +6% and Travel Time between +17% and +20% but reduce Fleet Size by 17% to 55%. The replacement of Conventional Public Transit (traditionally operated with high-capacity vehicles and with a line and schedule base) by AV or SAV could reduce Travel Time by half and might also help to reduce overall costs. This transition to on-demand services would also extend the required Fleet Size and the VKT by three to six times the initial value.
82Future studies could provide more accurate results from the mobility simulation of the type of territories or specific impact of one of the service features, such as ridesharing. In the meantime, the literature could benefit from an extended analysis of the impact of Autonomous vehicles through the socioeconomic prism.