Crash reports include details of the circumstances leading up to a collision, any resulting injuries and/or property damage, the geolocation of the collision, the type of crash, weather, roadway conditions, and time and date of the incident (Sinha et al., 2021; Novat et al., 2023). Disengagement reports provide information on every recorded disengagement, including road and weather conditions and other influencing factors (Dixit et al., 2016), and the open-source accessibility allows researchers to perform studies that use a variety of statistical models and descriptive analyses to examine AV collisions and disengagements and evaluate AV performance and safety concerns. Table 1 provides an overview of the literature reviewed.

As demonstrated in Table 1, multiple authors avow that rear-end and sideswipe collisions are the most frequently occurring types of crashes (Favarò et al., 2017; Leilabadi and Schmidt, 2019; Wang and Li, 2019; Boggs et al., 2020; Das et al., 2020; Petrović et al., 2020). Liu et al. (2021) utilized statistical analysis to identify the distinctions between the pre-crash scenarios of conventional cars and AVs and determined that the two groups differed in the number of collisions when the situation was the same. Based on the data points analyzed, AVs are 1.6 times more likely to be involved in a rear-ended collision than conventional vehicles due to AVs’ insufficient path planning or their inaccurate predictions of the behaviors of human drivers (Liu et al., 2021). The difference between the perception-reaction time of drivers of conventional vehicles and AVs may also play a role (Liu et al., 2021). Petrović et al. (2020) theorized that the frequency of rear-ended collisions may be attributed to the behavior of conventional vehicle drivers, as they are unaccustomed to the dynamic driving style of AVs that encompasses fully complying with traffic regulations and not being susceptible to human driving errors or behaviors. Thus, for most rear-end accidents, drivers of conventional vehicles are found to be at fault for driving too fast and too close to the AV.

Despite the increased likelihood of rear-end accidents, AVs are less likely than conventional vehicles to be involved in sideswipes, head-ons, and other types of collisions (Novat et al., 2023). Broadside and pedestrian collisions are significantly less likely to occur with AVs also, as they represent only about 6% of AV accidents compared to 42% of conventional vehicle accidents (Petrović et al., 2020). This difference can be attributed to AVs’ ability to better evaluate their surrounding environment and locate stationary or moving objects with the use of complex technological systems (Petrović et al., 2020).

Although it was determined that AVs are not responsible for 87% of the accidents they are involved in (Wang and Li, 2019), it is essential to identify the factors that contribute to the severity of the collisions that do occur. When a vehicle is involved in a collision, the severity level can range from no injury to a severe impact on the vehicle and persons involved, and in extreme cases can even result in fatalities (Yuan et al., 2022). It was determined through an analysis of CA DMV collision reports that most AV crashes result in minor damage to both the AV and the other vehicle, and most of the personal injuries involve back pain (Dixit et al., 2016). Back, head and neck injuries are the most common type of injury, with AV occupants representing 70.83% of those injured (Ye et al., 2021). As safety is considered the primary advantage of AVs, the factors associated with increased crash severity were identified and are depicted in Table 2.

While the CA DMV data illuminates how AVs operate in a mixed traffic environment, the findings may not accurately represent their performance due to limitations associated with the data. Specifically, the sample size of AVs considered in the published data is small compared with the number of conventional vehicles on the roads, and the data is limited to a specific geographical location that is not representative of all driving conditions in a global context. For instance, California does not experience adverse weather conditions, thus the findings do not illustrate how AVs perform in severe winter weather. Beyond that, it is unlikely that all driving conditions have been represented, as the driving environment is complex, dynamic, and unpredictable, and the limited amount of available published data on crash and disengagement reports make a comprehensive analysis challenging if not impossible. Furthermore, these findings cannot be generalized globally since the performance of testing vehicles is restricted to California’s roadway infrastructure and traffic regulations. Future research should account for this inadequacy by considering different locations and analyzing how additional factors may impact the operations of the technology.

AV interactions with other road users will likely receive more attention as the technology continues to develop. AV designers and manufacturers are continuously improving the vehicles’ capacity to screen their surroundings, detect nearby road users, and predict other drivers’ movements (Alozi and Hussein, 2022); however, despite their potential to increase road safety by minimizing human involvement, safety concerns are likely to arise if their interactions with other modes of transportation are not considered (Guo et al., 2022; Chen et al., 2023).

Interactions between road users can be unpredictable, even when all other factors remain constant (Chen et al., 2023), and AVs must be able to examine their surroundings and forecast the future behaviors of other road users before making control decisions (Parekh et al., 2022). This assessment process includes how the vehicle perceives its immediate surroundings, taking into account the infrastructure, obstructions, and nearby road users, as well as curbs, lanes, traffic signs, and signals (Hafeez et al., 2022). Radar, visible light cameras (VLC), and light detection and ranging (LiDAR) relay information about objects, traffic conditions, and lane occupancy to the driving system (Iftikhar et al., 2022; Parekh et al., 2022), but the foundation of AV technology is detecting elements of the road to ensure that it reacts appropriately to its surroundings and operates safely. The utilization of a multi-sensor system encompassing radar, VLC, and LiDAR is considered the optimum strategy since the advantages of one sensor offset the limitations of another (Combs et al., 2019). LiDAR uses scanning lasers to create a three-dimensional map that helps avoid collisions, while radar detection systems use radio waves to locate objects, and VLCs capture the surrounding area (Combs et al., 2019; Parekh et al., 2022).

Pedestrians are often regarded as the most vulnerable road users because of their frequent overrepresentation in traffic-related injuries and fatalities (Alozi and Hussein, 2022). Their actions and behaviors are highly unpredictable and probabilistic, they frequently appear in a variety of settings that are influenced by the surrounding environment, and the weather that can make it challenging to detect their presence (Combs et al., 2019) and distinguish between them and other visages in low visibility conditions and/or with low camera resolutions (Iftikhar et al., 2022; Parekh et al., 2022). Pedestrian-AV interactions are a crucial element in AV adoption and public trust (Kaur and Rampersad, 2018), but AVs have functional issues that make it difficult for them to detect pedestrians (Iftikhar et al., 2022). Fortunately, VLCs and a multi-sensor approach have the potential to change that and reduce pedestrian fatalities by approximately 90% (Combs et al., 2019).

Multiple researchers have conducted surveys that enabled them to evaluate public receptivity or attitudes towards AVs in relation to transportation safety (Pettigrew et al., 2019; Qu et al., 2019; Hasan et al., 2022), and their findings have consistently revealed that public perception is the biggest barrier to AV adoption, with safety being the primary concern. As a result, failure to ensure that vulnerable road users feel safe interacting with AVs could adversely affect other forms of transportation, such as walking and cycling, and hinder AVs’ integration into public roadways. Table 3 identifies strategies for policymakers and manufacturers that are designed to address AV safety concerns and their impacts on vulnerable road users.

Full utilization of AV technology is dependent upon manufacturers gaining the public’s trust and acceptance, and alleviating their concerns about scenarios involving inevitable collisions. The safety image of AVs has a considerable impact on how the public perceives autonomous driving technologies, and many studies have pondered who should be held legally and morally liable if a fully autonomous car causes a collision (Hevelke and Nida-Rümelin, 2015; Aguiar et al., 2022). In a potentially unavoidable accident, an AV has two choices, both of which have serious consequences: (1) stay on course and hit the approaching car; or (2) swerve and hit a pedestrian. This moral conundrum has been extensively discussed in the literature (Martinho et al., 2021), but the many potential outcomes and variables make it challenging to determine which choice is the most morally acceptable (Santoni de Sio, 2017). Autonomous cars take longer than human drivers in conventional cars to make pre-planned decisions that determine the vehicles’ actions (Robinson et al., 2022), and programmers need to consider how AVs react when faced with an unavoidable accident (Santoni de Sio, 2017). Table 4 identifies strategies for policymakers and manufacturers that are designed to reduce the moral dilemmas these unavoidable accidents present.

The ambiguous nature of ethics and individuals’ beliefs, values, opinions, and feelings make it challenging to determine what constitutes an AV’s most morally acceptable action (Martinho et al., 2021; Robinson et al., 2022), and everyone has his or her own set of personal ethical standards that can change, depending on a variety of things such as beliefs, values, opinions, and feelings. Because of this, manufacturers may make judgments based on experimental ethics, in which people are polled over the moral acceptability of various programming possibilities (Bonnefon et al., 2016); however, given the wide range of what is regarded as ethically acceptable or prohibited, this might not be a feasible approach (Santoni de Sio, 2017).

The public’s participation in the decision-making process is necessary to better understand their expectations of AVs’ behavior in unavoidable incidents and what decisions they consider ethical (Robinson et al., 2022). Several possible outcomes can be taken into consideration when an AV has an unavoidable collision, but since public acceptance is thought to be the main obstacle to the adoption of AVs, a widely acceptable solution should be put into place to increase public acceptance and trust. Because everyone has a different understanding of what is the proper course of action, the Massachusetts Institute of Technology conducted a study in which they polled individuals about the moral dilemmas involving AVs using an online experimental platform (Awad et al., 2018). Respondents from 233 nations provided 400 million decisions for the Moral Machine Experiment, which analyzed the preferred course of action that an AV should undertake, such as whether it should swerve, and examined case studies that contained sociodemographic characteristics of those injured and whether they were passengers or pedestrians (Awad et al., 2018). The study sought to quantify society’s expectations for moral behavior that should direct an AV’s decision-making process through the development of common ethical principles, but the findings emphasized the lack of homogeneity in what people view as moral. Thus, the validity of research like the Moral Machine Experiment has been questioned, as the results may be impacted by procedural or natural bias (Robinson et al., 2022).

Evans et al. (2020) avowed that utilizing a computational framework flexible enough to support a wide range of ethical standards is more effective than developing shared ethical principles since people have different views of the right moral course of action an AV should follow when involved in an unavoidable collision. They viewed AV decision-making as a form of claim mitigation in which people have various moral claims that influence how an AV reacts to both the user’s expectations and its environment. It is generally accepted that an autonomous driving system’s decision-making process is the key to reducing the number of crashes where human injury is likely to result. This has to do with the ability of the AV to simulate human judgment and driving behaviors while reducing human errors that are present in the traffic environment. Driving conditions are complicated, volatile, unpredictable, and include a wide range of diverse means of transportation; thus, potential solutions can be investigated by adopting virtual reality or driving simulation approaches (Robinson et al., 2022) that can enhance immersion and ensure process validity and reliability without compromising safety.

Guidelines, including the German Ethics Code for Automated and Connected Driving (Luetge, 2017) and the European Commission’s report on the Ethics of Connected and Automated Vehicles autonomous driving systems were developed in response to the ethical dilemmas presented by the scenarios of unavoidable collisions, where an AV has no choice but to be involved in a collision where human harm is a likely consequence (Bonnefon et al., 2020). These recommendations can be used to provide a framework for establishing an AV decision-making process that can be employed in the face of an inevitable collision, where an automated machine may have to decide who must live or die. Both guidelines cautioned against programming AVs to deliberately favor one life over another based on personal characteristics, such as prioritizing a child’s life at the expense of an adult’s. The situations elicit moral reflection and call for consideration of issues such as who bears responsibility when an AV runs over a person or whether an AV’s passengers can be held accountable when a crash occurs (Aguiar et al., 2022). Such factors affect how regulations and policies are introduced and have practical implications for stakeholders within the AV industry. Both the European Commission’s report and the German Ethics Code emphasize the importance of designing autonomous driving systems with effective collision avoidance and safeguards against inappropriate or unsafe operation by drivers of conventional vehicles (Luetge, 2017; Bonnefon et al., 2020). The moral dilemma will become less significant as the likelihood of an inevitable collision decreases (Robinson et al., 2022).

Questions regarding who should bear responsibility in the event of an inevitable collision are frequently at the center of discussions about AVs and ethical dilemmas. While putting all of the blame on the manufacturer seems like the obvious course of action, doing so could prevent further development of AVs (Marchant and Lindor, 2012). Due to the potential conflict of interest between the manufacturer’s liability concerns and the consumer’s desire for safety (Robinson et al., 2022), legislation should be created to structure tort liability for AV manufacturers that encourages the development and enhancement of AVs (Hevelke and Nida-Rümelin, 2015). Researchers have also explored the consumer’s role in collisions and whether AV users can be held accountable in the event of an unavoidable accident. While drivers have a duty to pay attention to traffic and the road and have the ability to foresee and prevent accidents by taking action when required (Hevelke and Nida-Rümelin, 2015), this solution has an exclusionary nature because it neglects users with restricted mobility who are dependent on fully autonomous vehicles for transportation.