The perils of perfect performance; considering the effects of introducing autonomous vehicles on rates of car vs cyclist conflict

Abstract

How humans will adapt and respond to the introduction of autonomous vehicles (AVs) is uncertain. This study used an agent-based model to explore how AVs, human-operated vehicles, and cyclists might interact based on the introduction of flawlessly performing AVs. Under two separate experimental conditions, results of experiment 1 showed that, despite no conflicts occurring between cyclists and AVs, modelled conflicts among human-operated cars and cyclists increased with the introduction of AVs due to cyclists’ adjusted expectations of the behaviour and capability of human-operated and autonomous cars. Similarly, when human-operated cars were replaced with AVs over time in experiment 2, cyclist conflict rates did not follow a linear reduction consistent with the replacement rate but decreased more slowly in the early stages of replacement before 50% substitution. It is concluded that, although flawlessly performing AVs might reduce total conflicts, the introduction of AVs into a transport system where humans adjust to the behaviour and risk presented by AVs could create new sources of error that offset some of AVs assumed safety benefits.

Practitioner summary: Ergonomics is an applied science that studies interactions between humans and other elements of a system, including non-human agents. Agent-Based Modelling (ABM) provides an approach for exploring dynamic and emergent interactions between agents. In this article, we demonstrate ABM through an analysis of how cyclists and pedestrians might interact with Autonomous Vehicles (AVs) in future road transport systems.

Abbreviations: ABM: agent-based model; AV: autonomous vehicle; ODD; overview, design concepts and details; RW: rescorla-wagner

Keywords:

• Cycling
• safety
• traffic
• health
• autonomous vehicles
• agent-based modelling

1. Introduction

The health benefits and reduction in chronic disease from increased physical activity gained through active transport are well established (Woodcock et al. 2011). However, in places where the poor separation between motorised vehicles and vulnerable road users exists, injury rates per kilometre travelled are comparatively higher than those where physical separation is present (Morrison et al. 2019; World Health Organization 2018). Worldwide, more than half of road traffic deaths occur among vulnerable road users who interact with motorised vehicles (World Health Organization 2018). It is therefore difficult for cities to encourage increased rates of cycling without experiencing concomitant (though not proportional) increases in cycling crashes (Bhalla et al. 2007; Stevenson et al. 2016). While investment in separated cycling infrastructure is one solution to this problem, high financial and political costs are often associated with these policies in already car-dominated countries such as the United States, Canada, Australia, and New Zealand, when they are perceived to divert funding and priority away from motor vehicles (Pucher, Dill, and Handy 2010; Reynolds et al. 2009).

As an alternative to costly infrastructure changes, the deployment of autonomous vehicles (AVs) has been forwarded as a potential solution to the issue of mixed-mode interaction (Combs et al. 2019; Millard-Ball 2018). If AVs can learn to accurately identify other vehicles and vulnerable road users and either evade, avoid or recover from situations leading to crashes, the benefits in terms of reduced injuries and trauma could be significant (Combs et al. 2019; Parkin et al. 2016). However, the prospect of coordination among networked AVs with vehicle-to-vehicle communication in controlled environments is very different to that of AVs interacting with non-networked and less predictable humans in situ – something Hancock (2019) refers to as mixed equipage. And while the idea of ‘flawless’ or perfectly performing AVs that do not crash exists, the technology required to ensure AVs can operate safely in the extreme variety of unpredictable and novel situations vehicles face, to ‘learn’ (in an algorithmic sense) from these experiences, and to then make acceptably ‘moral’ decisions in the case of inevitable collisions, is far from ready (Hancock 2018). Indeed, the introduction of AVs may lead to new types of crashes due to the pitfalls of automation (De Winter 2019).

Further, once deployed, community expectations hold that AVs will be safer than human drivers (Anderson et al. 2018; Fagnant and Kockelman 2014). In this context, the financial and legal consequences for AV manufacturers’ vehicles contributing to crashes may also be higher than for fallible human drivers (Anderson et al. 2018) and recovery outcomes for people injured as a result of AV errors potentially worse given resulting external attributions of responsibility for crashes (Pollanen et al. 2020; Thompson, Berk, et al. 2014; Thompson et al. 2015; Thompson, O’Donnell, et al. 2014).

Putting technological barriers aside, even if AV manufacturers achieve ‘flawlessly performing’ vehicles that never collide with humans, humans’ responses to the introduction of AVs is uncertain. A great strength of humans, and other intelligent animals (Legagneux and Ducatez 2013), is adaptability; we learn to understand, manipulate and exploit our environment. Consistent with other learned behaviours across alternate domains, if AVs behave differently around humans (i.e., to drive conservatively to avoid collisions at all costs), humans may learn to behave differently around AVs than they had previously around manually operated cars that bear the threat of injury (Brooks 2017; Millard-Ball 2018). This dynamic adaptation could lead to new situations and crash risks not previously present or imagined in the transport system. Whilst many pessimists have focussed on imperfect AVs and their safety impacts (Hancock 2018; Salmon 2019), few have considered the potentially adverse safety impacts of perfect automation within environments containing other imperfect actors. Indeed, recent detailed analyses of the potential of AV technology to prevent over 3300 recorded vehicle vs pedestrian crashes in the United States specifically notes the limitations of statistical models not being able to capture human and other drivers’ adaptation to the presence of AVs over the period of introduction (Combs et al. 2019).

There have been many models of traffic safety developed previously, which have taken on a variety of forms (Young et al. 2014). Overwhelmingly, these models have taken a static, top-down approach where the risk of a crash (and consequent crashes) is estimated using probabilities or risk gradients based on vehicle behaviours such as speed, angle, and the modelled physical proximity of vehicles to one another. Given that most of these efforts have developed from an engineering perspective, it is perhaps logical that many have been produced as extensions to existing traffic simulation models frequently used by traffic engineers designed for understanding traffic flow in applied situations (e.g. AIMSUN or VISSIM (Aimsun 2020; PTV Group 2020)). While efforts are regularly made to calibrate risk profiles to real-world data in these models (Davis and Morris 2009), the perceptual or skill errors leading to crashes that individual drivers experience remain black-box processes; we may know the risk of crashing between two vehicles crossing paths is x, but we still do not know why. Although game-theoretical models have addressed this gap to some extent by including static representations of cognitive frameworks in negotiated traffic conflict situations (e.g. Bjørnskau 2017; Elvik 2014; Millard-Ball 2018), to our knowledge, there are no traffic conflict or crash simulation models that approach the subject from the alternative perspective; that of dynamic individual cognitive and learning frameworks and manipulation of the factors that lead to perceptual errors that produce risky behaviours and consequent crashes.

The goal of this paper is, therefore, to better understand patterns of potential crashes (referred to in the model as ‘conflicts’) between vulnerable road users (namely cyclists) and manually-driven cars (cars) following the introduction of AVs into the transport system. We aim to explore changes in car-cyclist conflict frequency associated with the introduction of AVs into an existing transport system under conditions where 1) AVs perform ‘flawlessly’, 2) human drivers adapt to the presence of cyclists, and 3) cyclists adapt to the presence of AVs. The second goal of this paper is to demonstrate how this question can be tackled using methods of computational social science, namely, agent-based modelling. A third and final goal of the paper is to demonstrate the agent-based modelling method and its potential utility in assisting to answer questions in the field of ergonomics more broadly.

2. Method

This work used agent-based modelling, henceforth referred to as ABM. While the flexibility of ABM means that exceptions to any fast definition can likely be found, for the reader new to ABM, we would describe it in the following simplified manner: Agent-based modelling is a computational method that enables the exploration of interactions between multiple independent synthetic ‘agents’ or entities of interest. These agents generally make decisions and take actions in relation to a set of flexible rules governed by preferences, other agents’ observed behaviour, and in response to the potentially evolving nature of the environment they inhabit, which is also influenced by other agents in a bottom-up process. Although exogenous factors may also play a role in shaping the environment, the actions agents take influence their local environment. In turn, the environment has consequences for the agent, influences their state, and in turn, influences decisions and actions made by agents at subsequent time-points (see Figure 1). Agents can be modelled as independent entities at any scale; from individuals to populations, organisations or countries. Of interest in most ABMs related to socio-technical systems is the dynamic interplay between modelled agents at an individual scale and how this affects patterns of behaviour and performance of systems at a macro scale. Models are run dynamically over ‘time steps’ which can be used to represent any relevant time scale (i.e. from seconds to hours, to years or simply ‘moments’). For the first introduction to ABM, readers are referred to Gilbert (2008).

Figure 1. Depiction of the dynamic relationship between agents, their actions, the environment, consequences for the agent and consequences for the agents’ state.

The present model is an adaptation of a previously published (Thompson, Savino, and Stevenson 2016) ABM of behavioural adaptation by car drivers to the increased presence of cyclists represented within a simple, ‘Manhattan grid’ configuration consisting of 4356 unsignalised intersections, constructed in Netlogo (Wilensky 2016) (see Figure 2). A detailed description of the present model following principles adapted from the ‘ODD Protocol’ (Grimm et al. 2010) is available in Appendix A. In the prior model, drivers became increasingly aware of cyclists as more cyclists entered the road system, and, through an individual reinforcement learning and adaptation process represented by Rescorla–Wagner (RW) learning algorithms encoded within each agent (Rescorla and Wagner 1972), adapted their driving behaviour and likelihood of driving safely around cyclists in response. The RW learning model is representative of ‘classical conditioning’ whereby pairs of stimuli (e.g. roads and cyclists) are perceived together by an observer (e.g. a driver) to the extent that when presented with one stimulus (roads), the learner then expects the second, conditioned stimulus (e.g. cyclists), which may then trigger a conditioned response (e.g. more careful observance or driving behaviour). As with other representations of classical conditioning processes, observers also possess the capacity to have associations between stimuli extinguished if items are perceived independently.

Figure 2. Top-down, angled view of the Netlogo model showing the position of vehicles and vehicle classes represented in the system.

In this paper, we extended the representation from Thompson, Savino, and Stevenson (2016) to include an additional class of vehicle (AVs) by also providing cyclists with an ability to learn to recognise and adapt behaviour towards both human-driven cars (cars) and AVs. The process of adaptation for cyclists also followed an RW configuration, with each cyclist possessing an accuracy of perception, a desire to act well, a memory for previous events, a maximum level of association between stimuli, a timescale for building up knowledge, and a timescale for forgetting (i.e. a memory span).

2.1. Model context

An important distinction should be made for readers more familiar with traffic simulation models as are common within the engineering literature and typically devoted to crash risk estimation and safety (e.g. see Young et al. 2014). Firstly, expectations related to the accurate representation of spatio-temporal elements of the model should be put to one side. The model does not attempt accurate representation of time, speed, nor origin-destination matrices. Rather, it seeks to represent features related to learning and cognition in respect of interactions between vehicle operators of various classes; human and non-human. In this respect, it is more-so a model of learning and interaction applied to the transport domain than a traffic simulation model.

Secondly, the model does not attempt an accurate representation of traffic volumes but attempts to create a critical number of interactions between simulated vehicle operators that exposes only the variables of interest to analysis. Here, an analogy can be made with the well-known ‘El Farol Bar’ model (Arthur 1994) – one of the first to explore the outcomes of inductive reasoning and bounded rationality among computational agents. Within the El Farol Bar model, agents decide whether they will attend a bar based on how crowded they think it will be and this decision is based on their previous experience of attending that bar. The focus of the model is, therefore, the decision-making of agents and the macro patterns of behaviour their collective reasoning produces. Days of the week, time of day, hangovers experienced by the agents, or other features that might be used in models targeted for more pragmatic purposes (e.g. organising adequate bar-staff rosters) are ignored. Despite these exclusions, the model is instructive for its intended purposes in the tradition that it is designed to be both wrong and useful (Box 1976).

2.2. Model assumptions and scope

Assumptions made in the model were that AVs acted flawlessly by exhibiting extreme care, perfect identification of other vehicles, and always yielded to other human-driven cars and cyclists. AVs never collided with other vehicles. Human drivers acted identically to the manner they had done in Thompson, Savino, and Stevenson (2016); while desiring to drive as safely as they could around cyclists, their performance was occasionally imperfect, and their skill-levels changed in response to increasing or decreasing exposure to cyclists, consistent with the behavioural adaptation hypothesis (Jacobsen, Ragland, and Komanoff 2015). The mean number of conflicts for the population of drivers was equivalent to the inverse of the mean level of adaptation drivers had achieved at any time; more adaptation to cyclists, less conflict risk.

Cyclists’ adaptation in the model consisted of them learning that AVs would not crash into them. This was akin to cyclists developing schemata which influence their perceptual exploration and actions on the environment (Plant and Stanton 2013; Salmon et al. 2014). In the model, cyclist schemata incorporated flawless AVs that do not represent a safety risk; with increasing exposure, cyclists learned that they could safely progress through intersections without needing to yield to AVs because AVs operated perfectly. Like human drivers, however, cyclists’ perceptions were not without error; they did not always recognise the difference between manually driven cars and AVs and occasionally put themselves in the path of human drivers that did not have the same capacity to avoid a collision (i.e. perfectly safe performance) that AVs possessed. This capacity to recognise AVs in the road environment was represented through the value of saliency, described below.

Across the transport system, where cyclists encountered increasing numbers of AVs, an increased association between the act of cycling and the expectation of encountering an AV at an intersection occurred. Under these conditions, the association was expressed as a function of time, V(t_k).

Considering a given population of n cyclists, mean association to AVs at time t_k can be expressed as: (1) $$V\left(t_k\right)=\mathrm{ }\frac{\sum _{i=1}^n{V}_i\left(t_k\right)}{n}$$(1)

Where for an individual cyclist, i, the association is: (2) $$V_i\left(t_k\right)=V_i\left(t_{k-1}\right)+\Delta V_i\left(t_k\right)$$(2)

For the change in association $\Delta V_i\left(t_k\right)$ we assume a relationship of the form (3) $${\Delta V}_i\left(t_k\right)={{\alpha }_i\beta }_i{\omega }_i{\sigma }_i\left(\lambda -V_i\left(t_{k-1}\right)\right)$$(3)

In this equation, the values of ${\alpha }_i$ and ${\beta }_i$ represent the saliency of AVs that exist in the road environment and saliency of the road environment, respectively (i.e. how identifiable AVs are ($\alpha ),$ and how identifiable to the cyclist it is that the AV is on a road shared by them ($\beta$)). Changes to the value of these variables can be imagined as efforts to make AVs more or less distinguishable from ‘normal’ cars or making cyclists more or less visible (e.g. through the use of lights or high visibility clothing) to drivers. It is assumed that the value of road saliency for cyclists is 1 at all times, indicating that cyclists always recognise that the AV is on the road. The value of $\lambda$ represents the maximum possible association cyclists can achieve between cycling and expectation that cars are autonomous (AVs), and $V_i(t_{k-1})$ is the association strength between cycling and expectation of encountering an AV for the ith cyclist at the previous time step. In addition to these features, cyclists also held a fixed level of intention to ride safely around cars $({\omega }_i)$ and a capacity to do so $({\sigma }_i),$ which assumed a range between 0 (no intention or capacity to ride safely) and 1 (maximum intention and capacity) for each variable, respectively. These factors had the effect of moderating the impact of learned behaviour on risk-taking and therefore conflicts in a manner analogous to the final stages of the theory of planned behaviour (Ajzen 2011) (i.e. intention and actual behavioural control).

For cyclists in the model, the association between cycling and expectation of interacting with AVs remained stable for the period equivalent to a cyclist’s memory span, m. Cyclists’ memory spans were not infinite and the continued absence of AVs beyond a cyclist’s memory span resulted in a loss of association for individual cyclists. Such ‘extinction’ could occur at any stage in the learning process. Consistent with the Rescorla-Wagner model of conditioning, the decrease in association between cycling and expectation of interacting with an AV for individual cyclists also followed Eq. (3)(3) $${\Delta V}_i\left(t_k\right)={{\alpha }_i\beta }_i{\omega }_i{\sigma }_i\left(\lambda -V_i\left(t_{k-1}\right)\right)$$(3) , but with λ set to 0, reflecting the minimum possible association between ‘between cycling and expectation that cars are autonomous (i.e. no expectation of encountering an AV)’. This resulted in the final formulation of ${\Delta V}_i\left(t_k\right)$ as follows: (4) $${\Delta V}_i\left(t_k\right)=\left\{\begin{array}{l}{{\alpha }_i\beta }_i{\omega }_i{\sigma }_i\left(\lambda -V_i\left(t_{k-1}\right)\right)\\ -{\alpha \beta }_i{\omega }_i{\sigma }_i{V}_i\left(t_{k-1}\right)\\ 0\end{array}\mathrm{ }\begin{array}{l}\text{if at }\mathrm{least} 1\text{ AV }\mathrm{encountered}\text{ in }\left(t_{k-1},\mathrm{ }{t}_k\right);\\ \text{if no AV encountered in} \left(t_{k-m},\mathrm{ }{t}_k\right); \\ \text{ otherwise}.\mathrm{ }\end{array}\right.$$(4)

All other relationships in the model are consistent with those outlined in Thompson, Savino, and Stevenson (2016) (see Appendix A), including behavioural adaptation by human drivers to the presence of cyclists. Table 1 shows values associated with each agent class and characteristics within the model. In each trial, values for individual agent characteristics across vehicle saliency, capacity, and care were drawn from a normally-distributed range with a mean as indicated, and a standard deviation of 0.1. Where individual agent characteristics exceeded 1.0 for care, capacity and vehicle saliency, or was less than 0, the value was re-drawn.

Table 1. Factors, notation, and settings associated with each trial under each experimental condition.

Factor	Notation	Settings
Saliency (road)	$\alpha$	1.0
Saliency (AV to cyclist or bicycle to driver)	${\beta }_i$	0.6, 0.8, 1.0 (SD = 0.1)
Care	${\omega }_i$	0.6, 0.8, 1.0 (SD = 0.1)
Capacity	${\sigma }_i$	0.6, 0.8, 1.0 (SD = 0.1)
Memory span (time steps)	$m$	15, 30, 45

To test the effect that the introduction of AVs had on bicycle vs vehicle conflicts, the model was run using two experimental approaches. In the first, a ‘pre vs post’ comparison of conflict frequency was made between conditions consisting of a population of 2000 cars and 500 bicycles. The change in conflict frequencies was monitored before and after 500 additional AVs were deployed in the system at time-step 100 of 200. In the second approach, an additional 500 manually driven cars were added to the system at time 100 instead of AVs. The decision to add an additional 500 (25%) vehicles was taken to consider the effect of the policy change, only, and was not designed to reflect realistic assumptions.

Figure 3 uses pseudo-code to describe the interaction of cyclists among cars and AVs at intersections and the likelihood that cars crash into cyclists. Code block 1 shows the likelihood that cyclists would slow down and yield for vehicles at intersections is dependent upon their level of association (V) at that time point. Code block 2 demonstrates that if a cyclist and car find themselves at the same position in an intersection, the likelihood that conflict is recorded is dependent upon the driver’s level of adaptation to cyclists at that time point.

Figure 3. Pseudo-code associated with behaviour of cyclists (Block 1) and cars (Block 2) at intersections as it relates to crash-risk.

In the second experimental approach, an original population of 2000 manually driven cars was exchanged for AVs over 2000 time-steps. Cyclist numbers remained constant at 500 in each experiment. The values of factors contained within the RW formula for cyclists was varied across trials to determine how this contributed to changes in overall learning (V) among cyclists to AVs and consequent crash rates (i.e. from the RW learning model).

Table 1 shows the conditions under which each experimental condition was run, including variation in elements contained within the RW learning model described, above. In total, the model was run under a total of 243 separate conditions, 81 under the ‘pre vs post’ AV deployment approach, 81 under the pre vs post manual deployment approach, and 81 under the ‘trade’ approach, varying the strength of parameters described in Table 1.

3. Results

Results presented below represent mean statistics for the combination of all varied parameters across all model runs. Additional detail is also presented showing variation in model outputs taking into account differences between trials in agents’ memory span and learning rate as moderated by the saliency of (1) bicycles to drivers and, (2) of AVs to cyclists. We present these results separately to demonstrate the sensitivity of the model outputs to input parameters. Examples of how other researchers might want to tune model parameters for their own purposes could be where a researcher wishes to cap the level of saliency associated with a group of cyclists who do not use lights at night, or where the researcher wishes to model the increase in sensitivity to the presence of cyclists among drivers who have been primed to expect cyclists on the road through education campaigns (e.g. Dalton, Sumner, and Jones 2020). To allow for adequate model ‘burn-in’, pre vs post comparisons in experiment 1 are made between time-steps 50–100 (pre-condition) and 100–150 (post-condition).

3.1. Experiment 1: addition of 500 AVs to an existing population of vehicles vs 500 manually driven cars at time-step 100

Results of experiment 1 demonstrated that after the addition of 500 AVs at time-step 100, the mean number of conflicts in the simulated road system between manually-driven cars (cars) and cyclists increased from 54.73 (SD = 0.53) per time-step to 56.13 (SD = 0.92) (see Figure 4). Importantly, this increase occurred despite AVs operating ‘flawlessly’ and never colliding with any cyclists. Rather, the increased conflict frequency occurred between manually operated cars and modelled cyclists due to cyclists’ adjusted, imperfect expectations about the behaviour and capability of manually-operated and autonomous cars they were exposed to (see Figure 5(b) later in the text). Cyclists that had become accustomed to the safe behaviour of AVs occasionally mistook them for manually driven cars and expected them to operate as AVs did. Consequently, some cyclists failed to yield and were more likely to experience conflicts with drivers, particularly if the drivers had low levels of adaptation to cyclists. These effects were independent of the learning rate controlled by features such as vehicle saliency.

Figure 4. Mean count of car vs cyclist conflicts per time-step with additional deployment of AVs or Manual cars at time-step 100.

Figure 5. (a) Mean count of car vs cyclist conflicts per time-step with additional deployment of 500 AVs at time-step 100 under each of 3 AV saliency conditions, 0.6, 0.8, and 1.0. (b) Mean association between AVs and the road environment recorded for simulated cyclists per time-step with additional deployment of 500 AVs at time-step 100 under each of 3 AV saliency conditions, 0.6, 0.8, and 1.0.

By comparison, the introduction of 500 manually driven cars into the model at time-step 100 increased mean crash rates from 54.51 (SD = 0.41) to 62.94 (SD = 2.20). This result suggests that although the inclusion of AVs into the system increased overall conflicts, it did so to a considerably lesser extent than adding manually driven vehicles under our assumptions.

Closer inspection of these results reveals their sensitivity to model assumptions. Figure 6(a,b) show how the memory span of agents in the model affects total recorded conflicts pre and post-deployment. The effect of longer modelled memory spans (e.g. 45 time-steps) among agents is to dampen the impact of post-deployment bicycle vs car conflicts in the pre- vs post-conditions. This is because drivers with longer memories are less likely to forget having seen cyclists (i.e. levels of V remain elevated – see Appendix A). This effect outweighed the impact of cyclists’ increased memory for having observed AVs.

Figure 6. Mean count of car vs cyclist conflicts per time-step with additional deployment of (a) 500 AVs at time-step 100 under each of 3 memory span conditions of 15, 30, and 45 time-steps. (b) 500 manually-driven cars at time-step 100 under each of 3 memory span conditions of 15, 30, and 45 time-steps.

Comparison of results under varying conditions of cyclist saliency (i.e. β values associated with cyclists) also demonstrates the reduction in conflict frequency between manually driven cars and cyclists associated with greater cyclist saliency in both the pre-and post-deployment scenarios (see Figure 7(a,b)).

Figure 7. Mean count of car vs cyclist conflicts per time-step with additional deployment of (a) 500 AVs at time-step 100 under each of 3 bicycle saliency conditions, 0.6, 0.8, and 1.0. (b) 500 manually-driven cars at time-step 100 under each of 3 bicycle saliency conditions, 0.6, 0.8, and 1.0.

Finally, comparison of results under varying conditions of AV saliency (i.e. β values associated with AVs) across the pre-and post-deployment scenarios (see Figure 5(a)) showed little observed difference in conflict frequency between manually driven cars and cyclists under the three conditions (0.6 saliency M = 55.86, SD = 1.01, 0.8 saliency M = 56.21, SD = 1.13, 1.0 saliency M = 56.15, SD = 1.15). Figure 5(b) shows the mean strength of association (V) for cyclists across conditions in the post-deployment phase. This demonstrates that so long as learning occurred among cyclists and remained strong (i.e. V remained high) the rate of learning was less consequential for the frequency of conflicts, at least in the short-term post AV deployment. Interestingly, a lower (and therefore slower) learning rate ultimately produced a higher mean association level; a phenomenon familiar to many researchers in the fields of AI and machine learning where learning rates are tuned down toward the end of trials. Results for the manual-car deployment scenario are not shown as no AVs were deployed in that scenario.

3.2. Experiment 2: gradual introduction of AVs

Results of experimental condition 2, where manually driven cars were gradually traded for AVs over the course of 2000 time-steps, demonstrated another perspective. When manually driven cars were exchanged for AVs over time, cyclist conflicts did not follow a proportional linear reduction in frequency. We fitted a cubic spline with natural end-point conditions to the data points at time-step x = 0, 500, 1000, 1500 and 2000. Coefficients of the fitted spline function in the first segment [0, 500] show the function is concave, indicating the reduction in conflicts is non-linear and accelerates as more human drivers are replaced by AVs (see Figure 8). This phenomenon was caused by two functions of the model.

Figure 8. Count of car vs cyclist conflicts per time-step with trade of manually driven cars for AVs at a rate of 1 per time-step over 2000 total time-steps alongside a fitted cubic spline curve and risk of car vs cyclist conflicts per manually driven car.

Firstly, the initial non-linear reduction in conflicts per time-step and increase in crash-risk per time-step for manually driven cars was driven by the same behavioural adaptation function responsible for the increase in conflicts observed in experiment 1; cyclists that had adapted to AVs failed to yield to manually driven cars. This trend was most noticeable up to around time-step 500–1000 (25–50% of replacement) as confirmed by the fitted cubic spline curve. However, the rate of conflict reduction accelerated after this period alongside a reduction in risk of cyclist conflicts per manually driven car. This latter effect was driven by an increase in an average speed of bicycles throughout the network who were no longer stopping to give way to AVs as frequently, which consequently exposed remaining manually driven car drivers to more cyclists (see Figure 9). This resulted in a gradual increase in the level of association and behavioural adaptation to cyclists (and increased safety) among car drivers in the closing stages of replacement. While this result is consistent with results of prior models studying the behavioural adaptation hypothesis (Thompson, Savino, and Stevenson 2016) it is unclear whether it would hold in a real-world environment (see Section 4.4)

Figure 9. Mean speed of bicycles per time-step and mean association strength (V) of manually driven cars.

4. Discussion

The interaction between AVs and vulnerable road users is likely to represent a topic of considerable interest to road safety researchers and the travelling public over the next few decades (Hancock 2018; Kyriakidis et al. 2019; Salmon and Read 2019). Whilst most safety concerns relate to imperfect AVs that may not interact well with cyclists and pedestrians, other drivers or AV operators, we are unaware of other studies that have deliberately considered the negative safety impacts of well-designed and optimally functioning AVs in the context of interactions with imperfect humans outside the vehicle. In this study, we used ABM to explore potential road user behaviours and road safety outcomes during a transition period to flawlessly performing AVs under conditions of dynamic learning among vehicle classes.

The findings highlight that under conditions where humans (e.g. cyclists) adapt to AVs acting flawlessly, new sources of error could emerge that off-set the total safety benefits of AVs to some extent. The ABM demonstrated that even if AVs perceive and act perfectly, contributing no collisions from their own actions, if humans consequently adapt to a lower-risk environment created by their presence, it could increase the risk of conflicts between cyclists and manually driven cars during the transition. This could occur where vulnerable road users do not appropriately distinguish between AVs and manually driven cars. In particular, the modelling demonstrates that, under our assumptions, the introduction of additional AVs into a fleet of manually driven cars and cyclists of a static number could increase total cyclist conflicts due to increased risk-taking by cyclists. In an alternative ‘trade’ scenario that gradually replaces manually driven cars with AVs, we show that the reduction in total conflicts might not decrease proportionately with the introduction of perfectly performing AVs, but may decrease more slowly initially, before accelerating towards the end of the transition period (after 50% replacement). These results suggest that clear perceptual distinctions between AVs and manual vehicle classes may be required at the point of mass deployment to ensure safety benefits of AVs are not off-set by behavioural adaptation.

4.1. The influence of schemata on road user interactions

There is evidence in the ergonomics and safety science literature that aligns with the present study’s findings. In previous work examining the interactions between different road users, Salmon et al. (2014) demonstrated the critical role that road users’ previous experience and schemata play in creating safe and unsafe interactions between drivers and cyclists. In short, the experience of previous interactions with other road users has a strong influence on road users’ expectations and interactions with one another. For example, a cyclist’s schemata of how drivers behave at intersections direct where they look, what they perceive, and ultimately how they will behave. Accordingly, features of the schemata-based perceptual cycle model (Neisser 1976) and Norman’s (1981) schema-related error taxonomy are consistent with the model’s outputs. Based on repeated interactions with flawless AVs, cyclists will likely build AV schemata that create a strong expectation of flawless behaviour and low risk, which in turn will influence how they behave around AVs. Faulty activation of schema errors, specifically data driven activation errors (Norman 1981), whereby the AV schemata is triggered by manually driven cars which look similar to AVs, will result in cyclists behaving as they do around AVs. The paradox here is that, as AVs increase in number, cyclists’ ‘flawless AV’ schemata will be strengthened, and collisions with manually driven cars could become more likely as cyclists’ interactions with manually driven cars decrease. The schema explanation further explains adaptation by human drivers to the new cyclist behaviour in the second experimental approach; whereby drivers will be more likely to look for cyclists in the environment and take care to avoid collisions with them based on previous experience of their more risky behaviour around road vehicles (both AVs and manually driven cars).

4.2. AVs as unruly technologies

In discussion of the benefits of a transition to AVs, it is often reported that around 90% of road crashes are caused by human error (Infrastructure Victoria 2018; Parker et al. 1995). Perhaps concerningly, however, an implicit assumption underlying the use of this statistic for the justification of vehicle autonomy is that removal of humans from the driving task will reduce road crashes and trauma by a proportional extent (Fagnant and Kockelman 2014). This ignores an important element of accident causation across safety domains (e.g. patient, workplace & transport safety) that introduction of new technology designed for safety can also introduce new sources of risk and error of unknown severity and likelihood (Battles and Keyes 2002). These so-called ‘unruly technologies’ are a common contributory factor in major incidents (Dekker 2011; Grant et al. 2018). Moreover, there are previous examples of unruly technologies contributing to road trauma through the creation of new safety issues, including mobile phones and airbags (Salmon, McClure, and Stanton 2012). It is therefore critical that further studies are undertaken to explore the different emergent behaviours and risks associated with AVs. Whilst a focus on inadequate AVs is important (e.g. Hancock 2018), this study has highlighted that a focus on the potential adverse impacts of optimally functioning AVs is also warranted. This represents a significant gap in current AV research and further exploration is encouraged.

4.3. Agent-based modelling in ergonomics

By enabling the integration of a set of dynamic factors to be combined in a single computational model, the methods employed here demonstrate the utility of using ABM combined with learning algorithms to provide insight into emergent risk that would be difficult through discourse alone, or by using alternative modelling methods at the level of a functioning system. Indeed, while there is a long history of ergonomics studies in road safety (e.g. Brown 1979; Salmon and Read 2019; Stanton and Young 2005; Waard et al. 1995; Wagenaar and Reason 1990; Young et al. 2014) and more recent examples that have employed game-theoretic principles to explore interactions at non-signalised intersections (Bjørnskau 2017; Elvik and Bjørnskau 2017; Fox et al. 2018; Millard-Ball 2018) to date few have employed dynamic learning algorithms to explore the emergent behaviours associated with the introduction of new technologies or interventions as has been shown here. We hope this work highlights how ergonomics researchers and practitioners can explore and obtain new insights into phenomena through the combined use of ABM alongside existing models of learning and interaction both in road safety and in other ergonomics application areas.

4.4. Study limitations and future work

The limitations of this work are many and should be clearly noted. We explored results within a simulated, artificial transportation system where interaction between vehicle classes is restricted to simplified non-signalised intersections and the cognitive capabilities of agents are restricted to abstract (though well-established) learning principles. Human drivers’ adaptation to AVs is also not modelled, and no attempt has been made to realistically model travel demand, congestion, AV adoption patterns or other aspects that may come to mind for the reader familiar with traditional transportation modelling research. However, there are two responses to these limitations. Firstly, though what is left out of the model is substantial, what has been included is beyond what many models using traditional methods can integrate in terms of dynamic transition and behavioural adaptation among and between vehicles. Secondly, the ABM platform enables iterative improvement such that researchers interested in including omitted factors or relationships can do so, particularly because the code for the model is available for other researchers to access and utilise through an online repository (https://bit.ly/2YW76s9). In this respect, the model is open to the common criticism of many ABMs that it is at once ‘too complex’ and also ‘too simple’ (Waldherr and Wijermans 2013).

5. Conclusion

Humans are highly proficient at learning and adapting quickly to new environments, resulting in novel patterns of interaction. Regardless of whether AVs can learn to operate flawlessly in a transport system or not, humans will display new behaviours in response to AVs’ presence, especially when there are incentives (such as reducing travel time) to manipulate interactions. Such adaptation creates potentially novel crash situations and risks that are not typically in-scope of discussion when considering the safety benefits of AVs, which tend to focus on the role of vehicles or vehicle operators, only (Hancock 2018). Understanding how humans might adjust to future technological change in transport (and other domains) is challenging and requires methods that are both exploratory and empirical. We contend that dynamic modelling methods, including ABM, that can incorporate humans’ potential responses to AVs behaviour can be a highly useful method in the toolbox of ergonomics researchers.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

Funding Gemma Read’s contribution to this research was funded through her Australian Research Council (ARC) Discovery Early Career Research Award (DE180101449). Jason Thompson’s contribution was funded through his ARC Discovery Early Career Research Award (DE180101411). Paul Salmon’s contribution was funded through his ARC Future Fellowship (FT140100681).

Appendix A

Table

Adapted ODD Protocol (Grimm et al., 2010; Grimm, Polhill, & Touza, 2017)
ODD - Overview, Design concepts, and Details
Purpose of the model
The purpose of the model is to better understand potential patterns of crashes between vulnerable road users (namely cyclists) and cars following the introduction of AVs into a transport system. The model explores changes in car-cyclist crash frequency associated with the introduction of AVs into an existing transport system under conditions where AVs perform ‘flawlessly’, human drivers adapt to the presence of cyclists, and cyclists adapt to the presence of AVs.
Entities, state variables and scales
1. Manually driven cars (Drivers)
Manually driven cars possess states of speed, location, heading, memory span, perception for bicycles, perception for the roadway, levels of care to drive safely, capacity to drive safely, and association between cyclists and their presence on the road shared by the driver. Manually driven cars have capacity to perceive other vehicles located in their direct path.
2. Autonomous Vehicles (AVs)
AVs possess states of speed, location, and heading. They have the capacity to perceive other vehicles located in their direct path.
3. Cyclists
Cyclists possess states of speed, location, heading, memory span, perception for autonomous vehicles, perception of the roadway, perception of manual cars, level of care to ride safely, capacity to ride safely, association between autonomous vehicles and their presence on the road shared by the rider. Cyclists have capacity to perceive other vehicles located in their direct path and also perceive the number of other cyclists surrounding them in a 1 unit radius.
4. Roads
Roads are locations where vehicles travelled between intersections and can perceive one another but they cannot cross paths and cannot turn nor crash.
5. Intersections
Intersections are locations between roads where vehicles compete for space in order to continue their path. Crashes only occur at intersections.
Neither time nor distance in the model are scaled to represent real-world units.
Process overview and scheduling
Initialisation
At time step 0 (initialisation), the layout of the system is established including roads, intersections and placement of vehicles on empty roads and intersections in the network.
State variables associated with each vehicle class listed above are allocated to individual cyclists, manual cars and autonomous vehicles including initial location, speed, heading, association, and memory span.
Model Processing
Vehicle Class updates
At each time step, the following actions occur for members of each vehicle class:
All vehicles – Determine heading Update location Update speed Register presence of other at distance of 1 unit ahead Drivers – Calculate time since last observing a cyclist on the road Update level of association between cyclists and roadways Update likelihood of potential for collisions with cyclists given an interaction at current time-step Record collisions with cyclists Cyclists – Calculate time since last observing an AV on the road Update level of association between AVs and roadways Register presence of any manually driven cars with positive collision potentials at current location System Updates – Calculate total numbers of all vehicle classes in the system Calculate total numbers of collisions in the system Introduce new AVs into the system (Transfer condition – Experiment 2) Remove manually driven cars from the system (Transfer condition) If timestep = 250, deploy 500 AVs or manually driven cars into the system (Deployment condition – Experiment 1)
Output
Production of graphics and charts in user interface for interpretation Production of .csv datafile recording number of vehicles in each class, collisions, and levels of awareness among cyclists and drivers at each step for quantitative analysis
System Design	System Units
Area:
Length	130 units
Width	130 units
Area (individual patches)	16900 units
Intersections	4356
Agents:
Initial bicycles	500
Initial manually driven cars	2000
Initial autonomous vehicles	0
Speed per time-step:
Max speed of cars (manual and AV)	1.0 velocity unit
Max speed of bicycles	0.3 velocity units
Minimum speed of cars	0 velocity units
Minimum speed of bicycles	0.05 velocity units
Distance:
Distance units	1
Association:
Level of individual association among cyclists and drivers	Between 0 (minimum association) and 1 (maximum association)
Memory Span:
Number of time units each driver is able to remember observing a bicycle for and time each cyclist is able to remember seeing an AV for prior to association beginning to extinguish (m)	15, 30, or 45 time steps
Awareness:
Individual driver awareness of bicycles on the road (time since last observation of a cyclist < driver memory span)	Boolean (1 = aware, 0 = unaware)
Road saliency:
Extent to which drivers and cyclists consider paths travelled on by them to be ‘shared’ with cyclists and AVs, respectively (α)	Between 0 (not shared) and 1 (always shared) – set to 1 at all times
Bicycle Saliency:
Extent to which bicycles on the road are observed by drivers (β)	Between 0 (not observable at all) and 1.0 (always observed)
Care:
Individual drivers’ intention to drive safely among cyclists (ω)	Between 0 (no intention) and 1 (total intention)
Capacity:
Individual drivers’ capacity to drive safely around cyclists (θ)	Between 0 (no capacity) and 1 (total capacity)
Time:
Total time-step units elapsed (tk)	200 time units (Deployment condition)
	2000 time units (Transfer condition)
Change in vehicle class count per time step – Deployment condition:
Experiment a: Change in AV count	+500 at time-step 100
Experiment b: Change in manually driven car count	+500 at time-step 100
Change in vehicle class count per time step – Transfer condition:
Change in AV count	+ 1 per time step
Change in manual car count	−1 per time-step
Heading:
Direction of travel for vehicles	North, South, East, West
Exploration factor:
Likelihood that a vehicle of any class will change direction at any intersection	5%