PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract

Objective: This study investigated users’ subjective evaluation of three highly automated driving styles, in terms of comfort and naturalness, when negotiating a UK road in a high-fidelity, motion-based, driving simulator. Background: Comfort and naturalness are thought to play an important role in contributing to users’ acceptance and trust of automated vehicles (AVs), although not much is understood about the types of driving style which are considered comfortable or natural. Method: A driving simulator study, simulating roads with different road geometries and speed limits, was conducted. Twenty-four participants experienced three highly automated driving styles, two of which were recordings from human drivers, and the other was based on a machine learning (ML) algorithm, termed Defensive, Aggressive, and Turner respectively. Participants evaluated comfort or naturalness of each driving style, for each road segment, and completed a Sensation Seeking (SS) questionnaire, which assessed their risk-taking propensity. Results: Participants regarded human-like driving styles as more comfortable and natural, compared with the less human-like, ML-based, driving controller. However, between the two human-like controllers, only the Defensive style was considered comfortable, especially for the more challenging road environments. Differences in preference for controller by driver trait were also observed, with the Aggressive driving style evaluated as more natural by the high sensation seekers. Conclusion: Participants were able to distinguish between human- and machine-like AV controllers. A range of psychological concepts must be considered for the subjective evaluation of controllers. Application: Knowing how different driver groups evaluate automated vehicle controllers is important to design more acceptable systems in the future.
1
Drivers’ Evaluation of Different Automated Driving Styles: Is
1
It both Comfortable and Natural?
2
Chen Peng1*, Natasha Merat1, Richard Romano1, Foroogh Hajiseyedjavadi1, Evangelos Paschalidis1,
3
Chongfeng Wei1,2, Vishnu Radhakrishnan1, Albert Solernou1, Deborah Forster3, Erwin Boer3
4
5
6
1 Institute for Transport Studies, University of Leeds, 36-40 University Rd, Leeds LS2 9JT, UK
7
2 School of Mechanical and Aerospace Engineering, Queen's University Belfast, Belfast, BT7 1NN, UK
8
3 Entropy Control, Inc. San Francisco, CA 94107, USA
9
*Corresponding author. Address: Institute for Transport Studies, University of Leeds, LS2 9JT, UK.
10
Tel. +44(0)7925797354. E-mail address: C.Peng@leeds.ac.uk (C. Peng).
11
12
13
14
Précis
15
This study investigated the subjective evaluation of three automated driving styles, in terms of their
16
comfort and naturalness. Results showed a distinction between these two concepts, which is also
17
influenced by participants’ risk-taking personality trait, and the different road environmental factors
18
they experienced.
19
2
Abstract
20
Objective: This study investigated users’ subjective evaluation of three highly automated driving
21
styles, in terms of comfort and naturalness, when negotiating a UK road in a high-fidelity, motion-
22
based, driving simulator.
23
Background: Comfort and naturalness are thought to play an important role in contributing to
24
users’ acceptance and trust of automated vehicles (AVs), although not much is understood about
25
the types of driving style which are considered comfortable or natural.
26
Method: A driving simulator study, simulating roads with different road geometries and speed
27
limits, was conducted. Twenty-four participants experienced three highly automated driving styles,
28
two of which were recordings from human drivers, and the other was based on a machine learning
29
(ML) algorithm, termed Defensive, Aggressive, and Turner, respectively. Participants evaluated
30
comfort or naturalness of each driving style, for each road segment, and completed a Sensation
31
Seeking (SS) questionnaire, which assessed their risk-taking propensity.
32
Results: Participants regarded both human-like driving styles as more comfortable and natural,
33
compared with the less human-like, ML-based, driving controller. Particularly, between the two
34
human-like controllers, the Defensive style was considered more comfortable, especially for the
35
more challenging road environments. Differences in preference for controller by driver trait were
36
also observed, with the Aggressive driving style evaluated as more natural by the high sensation
37
seekers.
38
Conclusion: Participants were able to distinguish between human- and machine-like AV controllers.
39
A range of psychological concepts must be considered for the subjective evaluation of controllers.
40
3
Application: Insights into how different driver groups evaluate automated vehicle controllers are
41
important in designing more acceptable systems in the future.
42
Keywords: highly automated driving, driving style, comfort, naturalness, sensation seeking
43
4
1. Introduction
44
With higher SAE level AVs (SAE International, 2016), drivers will inevitably lose the controllability of
45
the vehicle, and the role of human drivers will shift from active controllers of the vehicle, towards
46
passive observers and passengers (Elbanhawi et al., 2015; Kaber & Endsley, 2004). There are several
47
subsequent concerns that might hinder the deployment of these vehicles, such as users experience
48
of comfort inside the AV (Elbanhawi et al., 2015). Comfort is crucial for an AVs implementation, as it
49
is found to be correlated with trust and acceptance (Paddeu et al., 2020; Siebert et al., 2013),
50
important elements for encouraging public uptake of these new forms of mobility (Madigan et al.,
51
2016).
52
Although there is currently no commonly agreed definition for comfort in this context, some
53
suggestions exist. Under the context of automated driving, Hartwich et al. (2018) summarised driving
54
comfort as ‘a subjective, pleasant state of relaxation given by confidence and an apparently safe
55
vehicle operation, which is achieved by the removal or absence of uneasiness and distress’ (p. 1019).
56
For automated vehicles, however, comfort is not simply limited to physical aspects of the vehicle,
57
such as good seat design (Ebe & Griffin, 2001), or acceptable levels of engine noise, and vehicle
58
vibrations (Qatu, 2012). These features are mentioned in studies of traditional, manually operated,
59
road vehicles, and also in other domains, for example, cabin noise in aircraft (Pennig et al., 2012).
60
Since the vehicle is no longer controlled by a human, it is important that its “driving behaviour”, and
61
how it negotiates different road geometries, and traffic conditions, is considered pleasant, and rated
62
positively by the user, ensuring it feels comfortable and safe (Elbanhawi et al., 2015; Summala, 2007).
63
Other, more psychological, terms and concepts used in this context include ensuring the AV is
64
considered reliable, and familiar, avoiding any sudden surprise behaviours, which are shown to
65
5
enhance the acceptance, satisfaction and perceived safety of AVs (Carsten & Martens, 2018; Ramm
66
et al., 2014).
67
One, relatively unexplored, concept in this context is “naturalnessof the AV’s driving behaviour,
68
which has been linked to the familiarity of the AV’s manoeuvres, for the user. Here, the familiarity of
69
AV movements, rendered by mimicking human-like vehicle controls, is expected to fulfil human users’
70
anticipation of an AVs behaviours, and result in positive subjective feedback (Butakov & Ioannou,
71
2015; Hartwich et al., 2018). Moreover, Elbanhawi et al. (2015) suggest that naturalness of
72
automated driving is an important determinant of comfort. However, some empirical studies have
73
shown that familiar automated driving manoeuvres do not always lead to higher subjective comfort
74
(Hartwich et al., 2018), which suggests that more knowledge is needed on the link between these
75
two concepts, since they will likely contribute to acceptance of future AVs.
76
From a technical perspective, there are a large number of automated driving styles that could be
77
generated for such investigations. Taking motion planning as an example, the generated driving
78
behaviour of AVs could be robotic, with algorithm-optimised trajectories, based solely on sensory
79
information provided by lasers, radars and cameras, to adapt to the environment (e.g., Urmson et al.,
80
2008). Alternatively, these may mimic a human driver’s average behavioural patterns, by training
81
models, based on real human driving data (e.g., Hajiseyedjavadi et al., submitted; Rehder et al., 2017;
82
Wei et al., 2019). Personalisation of driving styles can also be achieved by using usersown driving
83
style in the model development loop (e.g., Menner et al., 2019).
84
Studies on manual driving suggest that participants’ reported levels of comfort are also linked to the
85
vehicle’s “driving style (Ellinghaus & Schlag, 2001, cited in Bellem et al., 2018), which is defined as
86
the driving habits of the driver, such as their preferred speed, threshold for overtaking, headway
87
distance, and tendency to violate traffic regulations (Elander et al., 1993). In highly automated
88
6
vehicles, the use of such driving styles has been reported to enhance driving comfort of passive users
89
(Bellem et al., 2018).
90
Research has revealed the existence of several driving styles, associated with different character
91
traits of human drivers, loosely linked to defensive (less sudden acceleration and deceleration) and
92
aggressive (higher acceleration and more sudden braking) driving behaviours (Murphey et al., 2009).
93
Results also suggest that different automated driving styles are sometimes found to be preferred by
94
different groups of users, when evaluated in terms of comfort, safety and pleasantness, although
95
findings are inconclusive. For example, a more defensive driving style, with slower lane changing
96
features, and lower acceleration, was favoured by most participants, when compared with a higher-
97
acceleration, more assertive, driving style (Rossner & Bullinger, 2020). Moreover, Hartwich et al.
98
(2018) found that familiar driving styles (a replay of participants’ own driving) were more favoured
99
by younger drivers (25-35 years), while faster and unfamiliar automated driving styles (that of the
100
younger drivers) were preferred by older drivers (65-84 years). Therefore, Hartwich et al. (2018)
101
suggest that solely mimicking drivers personal manual driving habits may not be suitable for all age
102
groups. Using a more comprehensive set of vehicle kinematics, Bellem et al. (2018) manipulated the
103
initiation time and strength of acceleration and jerk of three manoeuvres on the highway (i.e., lane
104
changes, accelerations and decelerations). These authors recommend a number of configurations for
105
comfortable driving experiences, such as minimising jerk for acceleration and deceleration
106
manoeuvres, lowering acceleration, and providing action feedback, which is when maximum
107
acceleration is applied at the early stages of a lane change manoeuvre.
108
As outlined above, most of the existing studies considering users’ responses to different driving styles
109
of AVs have compared different replays of drivers’ manual driving performance. To date, there has
110
been little comparison of user preferences for machine- versus human-like AV driving styles. An
111
7
important consideration here is the balance between what is expected from users about the
112
acceptable driving style of an AV, compared to that of a human driver. For example, studies have
113
shown that an AV controller that precisely follows the lane centre, is considered more competent,
114
compared to those with less accurate lane-tracking and more lateral drifts from the centre lane (Price
115
et al., 2016). Therefore, from a human factors perspective, more research is warranted to understand
116
what types of driving styles and behaviours of machine- and human-like driving are considered more
117
comfortable and natural, and whether these are linked to the particular driving environment being
118
negotiated by the AV.
119
Users’ perception of an AV’s driving style is known to be influenced by both objective and subjective
120
factors. For example, road furniture and geometry are known to influence ratings of safety and
121
comfort (Hajiseyedjavadi et al., submitted) and physiological response (Beggiato et al., 2019;
122
Radhakrishnan et al., 2020), while a number of studies have shown a correlation between personality
123
traits such as Sensation Seeking (Arnett, 1994) and preferred driving style. For example, in manual
124
driving; drivers with high sensation seeking scores are found to drive in a riskier and more aggressive
125
manner and at higher speeds, while low sensation seekers have a tendency to drive more slowly
126
(Louw et al., 2019; Taubman-Ben-Ari et al., 2004; Zuckerman & Neeb, 1980). However, results are
127
mixed regarding preferences for AV-driving styles. For example, Yusof et al. (2016) reported that both
128
assertive and defensive drivers, characterised by higher and lower sensation seeking scores,
129
respectively, showed a consistent preference for a defensive (and not assertive) AV driving style.
130
Therefore, in addition to considering user response to two human-like and one machine-like AV
131
controller, this study assessed the effect of road geometries and users’ sensation seeking scores on
132
such evaluations.
133
8
Current study
134
This study is based on data collected from a driving simulator study within the UK-funded
135
HumanDrive project (TS/P012035/1); the main purpose of which was to develop, and evaluate,
136
advanced AV-controllers, imitating natural, human-like, driving styles. Two representative human-
137
like driving styles were recorded, and replayed to participants. Response to these was compared to
138
a machine-like, machine learning (ML)-built, driving style (Solernou et al., 2020 ).
139
The following research questions were addressed in the present study:
140
1. Are the three driving styles rated differently in terms of perceived comfort and naturalness?
141
2. Do environmental settings influence the comfort and naturalness of the three driving styles?
142
3. Do users’ sensation seeking propensities affect their ratings of comfort and naturalness
143
towards the three driving styles?
144
4. Is a natural driving style also a comfortable driving style?
145
2. Methods
146
Participants
147
Twenty-four participants (12 male, 12 female) aged between 20 and 49 years (M = 35.7, SD = 7.1)
148
were recruited. All participants held a valid UK driving licence, with experience ranging from 2 to 27
149
years (M = 14.7, SD = 7.8). Reported annual driving distance ranged from 500 to 18000 miles (M =
150
7554.2, SD = 3982.7).
151
All participants were recruited by using the University of Leeds Driving Simulator database, and all
152
provided informed consent to take part in the study. Each participant was compensated £30 for
153
9
taking part in the study. This study was approved by the University of Leeds Ethics Committee
154
(LTTRAN-086).
155
Apparatus
156
The high-fidelity, motion-based University of Leeds Driving Simulator (UoLDS) was used in the
157
experiment. The simulator’s vehicle cab is based around a 2006 Jaguar S-type, housed within a 4m
158
diameter, spherical projection dome. There are eight visual channels rendered at 60 frames/s,
159
predominantly at a resolution of 1920×1200, providing a horizontal forward field of view of 270°.
160
The simulator also incorporates an eight degree-of-freedom electrical motion system. The
161
generated range of acceleration of the motion system is ±5.0 m/s2 (Jamson et al., 2007)
162
Experimental Design
163
A fully within-participant experimental design was used in this study to investigate participants’
164
subjective evaluation of three different automated vehicle driving styles, described below.
165
Participants were asked to use an eleven-point, Likert-type, scale, to rate how “comfortable” and
166
“natural” each automated drive felt, as it negotiated the same stretch of road, in six separate drives,
167
completed over two days.
168
2.3.1. Driving styles
169
A machine learning (ML) based controller, and two human-driven controllers were developed for
170
evaluation in this study. These controllers are described further below, and a diagram presenting the
171
development procedure is shown in Figure 1.
172
10
173
Figure 1. Overview of the development of the three AV controllers
174
The ML-built controller (Turner)
175
The ML controller was calibrated using a Recurrent Convolutional Neural Network (RCNN) that was
176
capable of imitating the human driving behaviour, in terms of future yaw rate and speed demands.
177
The RCNN was trained from data of 10 participants, from an earlier experiment of the project (see
178
Solernou et al., 2020). This controller will be called the Turner controller from here on.
179
The human-like controllers (Termed Aggressive and Defensive controllers)
180
The two human-like controllers were recorded drives of human participants, collected before the
181
main study took place, which were then replayed to participants of this study. Previous studies have
182
shown a positive correlation between speed choice and sensation seeking (Louw et al., 2019) as well
183
as risk-taking behaviour in manual driving (Ge et al., 2014; Oppenheim et al., 2016; Riendeau et al.,
184
11
2018; Ulleberg & Rundmo, 2003). To ensure that distinct differences in driving behaviour would arise
185
between the two human-driven controllers, recruitment of participants used for the human-driven
186
controllers was based on their sensation seeking scores.
187
Before recruiting participants for these replay drives, data from a previous study of the project was
188
used to create clusters of driving behaviour (see Appendix 1). These participants were clustered into
189
three main groups: defensive, moderate and aggressive drivers. There was a moderate, but
190
insignificant, correlation between participants’ sensation seeking scores, and cluster membership
191
(r(14) = .429, p = .143). For example, we found that the aggressive driving cluster contained
192
participants with higher sensation seeking scores. The absence of a significant correlation was likely
193
due to the small sample size used in this study. Following this analysis, participants with higher
194
sensation seeking scores from the aggressive cluster, and lower sensation seeking scores from the
195
defensive cluster were contacted to participate in the replay recordings of the current study. In total,
196
eight participants were recruited, four for each sensation seeking group (Table 1).
197
Table 1 - Descriptive statistics of the participants used for the replay recording phase
198
Gender
Age
Male
Female
Mean
Std.
High sensation seeking
4
0
36.25
9.78
Low sensation seeking
2
2
52
6.73
Note. AISS scores were calculated based on drivers’ responses to Arnett Inventory of Sensation Seeking (AISS; Arnett,
199
1994), and scores were the sum of all responses to in total of 20 questions, with a higher score means higher sensation
200
seeking propensities.
201
During the recording process, each participant drove the experimental route three times. The process
202
took approximately one hour. After the data collection, the clustering process was applied again for
203
the new data, to confirm the obtained driving behaviours belonged to the previously identified
204
12
defensive and aggressive driver groups, respectively. Out of the eight participants recorded, the
205
manual driving data of two participants (one per sensation seeking group, with scores of 59 and 43,
206
respectively) which was closest to the median of the defensive and aggressive clusters, were selected
207
as the representative driving styles for our two human-like controllers. It is worth highlighting that
208
the selected drives were also checked to ensure that no unusual or unexpected manoeuvres existed
209
along the drive. For the rest of this paper, the higher sensation seekers driving style will be termed
210
Aggressive, and the lower sensation seekers driving style will be called the Defensive driving style.
211
2.3.2. Road Environment and Scenarios
212
The simulated driving scene was modelled from real stretches of road around North Bedfordshire in
213
the UK (Figure 2). Two loops, going North and South, were simulated, creating a virtual environment
214
covering around 12 miles of driving. In the present work, however, only the North loop was included
215
for the simulated drive, since it included the range of scenarios required for studying driver behaviour
216
in response to changes in speed and geometry, and shortened the overall drive. This section of road
217
was approximately 5 miles long, taking about 15 minutes to complete.
218
To understand user preferences for, and in response to, a wide range of road geometries and speed
219
profiles, the layout of the North loop contained a combination of high-speed (60 mph) rural sections,
220
with varying road curvature, and more built-up, village sections, at a speed limit of 40 mph (Table 2).
221
13
222
223
Figure 2. Example of the simulated (top) and real (bottom) road environments
224
14
Table 2 - The speed limit and geometrical details of the simulated road.
225
Zone
Curve radius
Curve direction
Road type
Speed limit (mph)
Road context
1
300-800m
Left
Rural
60
Kerb + grass and the bridge in the middle of the area
2
Straight
Straight
Rural
60
Kerb + grass with hedge far from the road edge
3
<150m
Left
Rural
60
Kerb + grass and trees far from the road edge
4
<150m
Left
Rural
60
Kerb + grass with hedge quite far from the road edge
5
300-800m
Right
Rural
60
Kerb + grass with hedge quite far from the road edge
6
<150m
Right
Village
40
Kerb + grass and some structures far from the road edge
7
<150m
Left
Village
40
Kerb + grass
8
200-300m
Right
Village
40
Kerb + grass and hedge around 1-2m from the road edge
9
Straight
Straight
Village
40
Kerb + grass and fence quite close to the road edge
10
150-200m
Right
Village
30
Kerb + pavement and village structures far from the road edge
11
150-200m
Right
Village
30
Parked cars zone
12
300-800m
Right
Village
30
Kerb + grass and village structures
13
200-300m
Left
Rural
40
Grass, bushes and trees not close to the road edge
14
300-800m
Right
Rural
40
Grass and hedge far from the road edge
15
<150m
Right
Rural
60
Grass and trees far from the road edge
16
<150m
Left
Rural
60
Hedge at the road edge
17
150-200m
Right
Rural
60
Grass and hedge far from the road edge
18
300-800m
Left
Rural
60
Grass and bushes around 2m from the road edge
19
Straight
Straight
Rural
60
Fence around 1-2m from the road edge
20
Straight
Straight
Rural
60
Hedge at the road edge and an intersection at the end of the section
21
na
na
University
30
Mini roundabout and road markings
22
<150m
Left
University
30
Parked cars zone
23
300-800m
Left
University
30
Kerb + pavement
24
Straight
Straight
University
30
Kerb + pavement
226
15
Variables
227
The dependent variables were comfort and naturalness of the driving experience, for each controller.
228
A search of the literature at the time of study design revealed an absence of a formal, and universally
229
agreed, description for the two terms. To ensure that the same term was understood by all
230
participants, we therefore used a small expert group within the project team to define the two terms,
231
and included this information in the participant briefing sheet:
232
i. Comfortable driving was defined as ‘a driving style that does not cause any feeling of
233
uneasiness or discomfort;
234
ii. Natural driving was defined as ‘a driving style that is closest to your own driving’.
235
Participants evaluated each controller, in two ways: (i) after each drive, participants were asked to
236
provide an overall rating, based on their entire driving experience., and (ii) throughout the drive,
237
immediately after they heard a short auditory beep, which was played via the car’s speakers,
238
corresponding to 24 relevant sections in the drive (Table 2). They were taught to use a Likert-type
239
scale for guiding their responses, providing a number between -5 (Extremely
240
Uncomfortable/Unnatural) and +5 (Extremely Comfortable/Natural) (Figure 3).
241
Participants also completed the Arnett Inventory of Sensation Seeking questionnaire (Arnett, 1994)
242
after they finished the last drive. This questionnaire includes twenty items, and four response options
243
for each item, ranging from 1 (does not describe me at all) to 4 (describes me very well). Reverse-
244
worded items were further reverse-coded. We used the sum score of these items to characterise
245
sensation seeking tendency, with a higher score indicating a higher sensation seeking tendency.
246
16
247
Figure 3. The Comfort and Naturalness scales used in the study
248
Procedure
249
To reduce the effect of fatigue on participants, the study was conducted over two separate days
250
(M=6.75 days apart, SD=2.17), with data collection lasting about 1.5 hours on each day. Participants
251
evaluated the three driving styles twice: once in terms of comfort, and once in terms of naturalness,
252
with half of the participants evaluating in terms of comfort on day 1, and the other half on day 2,
253
(Figure 4).
254
Upon arrival on the first day, each participant received a written and verbal briefing of the study,
255
including how to use the subjective scale, and provided their written consent to take part in the
256
experiment. They then started the simulator experiment with a practice drive in manual driving,
257
followed by a practice ride in an automated driving mode. A researcher accompanied participants
258
17
during the practice session, which lasted 20-30 minutes. Following the practice drive, the researcher
259
left the simulator dome, and the participant started the first of three experimental drives, one for
260
each controller. The order of the three automated driving styles was counterbalanced across
261
participants, and participants left the simulator dome after each drive, to reduce fatigue effects. After
262
the second day’s experiment, participants were asked to complete a set of questionnaires, including
263
the sensation seeking questionnaire. The data from the other questionnaires is not reported here.
264
265
Figure 4. The overall experimental procedure, including the order of drives.
266
3. Results
267
The main aim of the analyses was to assess users’ evaluation of the three automated controllers, in
268
18
terms of comfort and naturalness. Participants’ subjective feedback about the driving styles adopted
269
by the controllers was provided in two ways: (i) an overall evaluation of the controller, after finishing
270
the entire drive, and (ii) 24 responses, based on the 24 auditory beeps throughout each drive, which
271
prompted a response for each of the different driving zones. Two statistical tests were used: the
272
Wilcoxon signed-rank test was used for the overall evaluation provided at the end of each drive, and
273
the Generalised Estimating Equation (GEE) (Liang & Zeger, 1986) was used for the 24 individual
274
evaluations provided during the drive (see configurations of GEE in Appendix 2).
275
Subjective evaluations of the driving styles
276
Table 3 shows results of the Wilcoxon signed-ranks test on matched-pairs, and Figure 5 shows box
277
plots of overall comfort and naturalness evaluation of the three driving styles.
278
Regarding overall comfort, the Wilcoxon test showed significantly higher evaluation for the Defensive
279
controller, compared to the Aggressive and the Turner controllers (both p <.001). There was no
280
significant difference between the Aggressive and Turner controllers (p = .49). Regarding overall
281
naturalness, the Wilcoxon test showed a significantly higher evaluation for the Defensive controller,
282
compared to the Turner (p < .001), and a higher evaluation than the Aggressive controller (p = .02).
283
A significantly lower score for the Turner than the Aggressive controller (p = .01) was also revealed.
284
Figure 5 shows that the evaluation of the Defensive driving style was relatively consistent across
285
participants. By contrast, the evaluation for the Aggressive and Turner controllers was more variable,
286
with a bimodal pattern observed in response to naturalness of the Aggressive, and the comfort and
287
naturalness of the Turner. To understand this further, we conducted additional analyses by taking
288
participants’ personality trait into account.
289
19
Table 3 - Wilcoxon signed-rank test results for overall comfort and naturalness
290
Comfort (overall)
Driving style
z
p
r
Defensive vs Turner
4.27
0.000*
0.87
Defensive vs Aggressive
4.11
0.000*
0.84
Aggressive vs Turner
0.70
0.490
0.14
Naturalness (overall)
Driving style
z
p
r
Defensive vs Turner
3.67
0.001*
0.75
Aggressive vs Turner
2.44
0.010*
0.50
Defensive vs Aggressive
2.25
0.020*
0.46
Note. * p < .05. Orders of paired comparison are based on z values.
291
292
Figure 5. Evaluation of each controllers’ driving style, in terms of its overall comfort (left) and overall
293
naturalness (right). Horizontal lines inside each box represent the median values. Whiskers denote a
294
distance of 1.5 times interquartile range (IQR) above the upper quantile up to the largest observation,
295
or below the lower quartiles up to the smallest value. Grey dots represent data points (with small
296
variations added to the position to avoid overlapping), while black points represent outliers.
297
298
20
Table 4 presents the GEE results for the repeated ratings of comfort and naturalness. Regarding
299
comfort, the probability of reporting high levels of comfort was significantly higher for the Defensive
300
controller, compared to both the Turner (OR = 7.21, p < .001), and Aggressive controllers (OR = 4.01,
301
p < .001). Comfort ratings were also slightly higher for the Aggressive, than the Turner, controller (OR
302
= 1.80, p = .04). Regarding naturalness, both the Defensive, and Aggressive controllers were more
303
likely to be rated more natural, when compared to the Turner controller (OR = 4.98, p < .001; OR =
304
2.59, p = .002). The Defensive controller also had a higher probability of being assessed as more
305
natural than the Aggressive driving style (OR = 1.92, p = .01).
306
Table 4 - GEE model parameter estimates and odds ratios for repeatedly reported comfort and
307
naturalness
308
Comfort
Driving style
Coefficient
SE
Wald
Sig
Odds ratio (OR)
Defensive vs Turner
1.975
0.238
68.847
0.000*
7.206
Defensive vs Aggressive
1.388
0.234
35.111
0.000*
4.007
Aggressive vs Turner
0.587
0.285
4.258
0.039*
1.799
Naturalness
Driving style
Coefficient
SE
Wald
Sig
Odds ratio (OR)
Defensive vs Turner
1.606
0.232
48.113
0.000*
4.980
Aggressive vs Turner
0.953
0.305
9.745
0.002*
2.593
Defensive vs Aggressive
0.653
0.255
6.560
0.010*
1.921
Note. * p < .05. Orders of paired comparison are based on the odds ratios.
309
To further understand whether subjective evaluation was due to any differences in the driving styles
310
of the controllers, the vehicle kinematics, including the speed and lateral offsets of all three
311
controllers, were inspected (Figure 6 and Figure 7). The interpretations provided in this section were
312
based on visual observations of the controllers’ kinematic characteristics only, and no formal analyses
313
were conducted. Figure 6 shows that, overall, speed was higher in the Aggressive driving style,
314
compared to the other two controllers. The Defensive and Turner controllers had similar increasing
315
21
or decreasing trends in speed, for the same road sections, with smoother patterns (i.e., less frequent
316
and gentler fluctuations in speed) seen for the Defensive controller. There was not much difference
317
in the observed lateral offset of the three controllers (Figure 7).
318
319
Figure 6. The speed profiles of the controllers
320
321
Figure 7. The lateral offset profiles of the controllers
322
22
The effect of road environment on subjective evaluation
323
As shown in Table 2, the simulated road included a range of road environments divided into five main
324
categories: (i) road type, (ii) speed limit, (iii) road context, (iv) curve direction, and (v) curve radius.
325
For simplicity, only two categories were included in this analysis, as follows:
326
i. road type (rural and village), which differed by posted speed limit (60 mph vs 40 mph), and
327
roadside furniture (see Table 2);
328
ii. curve radius (five levels, varying from straight sections to curves of less than 150m).
329
Road type was included as the representative of speed limit and road context, supported by the
330
strong (r = 0.88, p < .001) and medium (r = -0.05, p < .001) correlation between road type and the
331
two categories (speed limit, road context), respectively. The University road type was excluded from
332
analysis, due to the small number of sections falling into this category. The direction of a curve was
333
also not included as a factor, as it was not expected to have a significant influence on results. It is
334
worth mentioning that the number of road sections in each level outlined above was not equal, since
335
the road was a replication of the real world.
336
Figure 8 shows the average comfort and naturalness ratings for the three driving styles, for the
337
different road sections. An overall pattern was observed, such that, with increasing curve radius,
338
there was a mild reduction in both comfort and naturalness ratings for all controllers, especially in
339
the Rural areas. This pattern was not apparent in the Village areas, apart from two unexpected
340
fluctuations. Inspection of the vehicle-based metrics showed a high speed for the Aggressive
341
controller in the 200-300 Curve Radius section, and a suddenly changing speed of the Turner in the
342
300-800 Curve Radius section (for further evaluation of these, see Appendix 3).
343
23
344
345
346
Figure 8. Mean evaluation scores for comfort (left) and naturalness (right) for each driving style, for
347
the different road environments. Error bars indicate standard error of the data.
348
Table 5 shows the results of the GEE models, which showed that the effect of driving style, on comfort
349
and naturalness ratings, was significant for both the Rural and Village road sections (all p < .001). In
350
the Rural sections, which had a generally higher speed limit than the Village sections, there was less
351
difference in odds ratios between the Aggressive and the Turner controllers for the gentler roads (i.e.,
352
Straight and Curve Radius 300-800), but this difference was more prominent for the shaper road
353
sections (i.e., Curve Radius 150-200, and 200-300). It is also worth highlighting the preference for the
354
Defensive controller over the Turner, where the odds ratios are seen to be larger with increasing road
355
curvatures. However, these differences in controllers were not observed for the sharpest Rural
356
24
section (i.e., less than 150).
357
In the Village sections, where the controllers negotiated the road at a lower speed, the observed
358
pattern with curvature outlined above, was not as apparent. This may be because all controllers
359
negotiated the curves at a relatively low speed, thus reducing the effect of Curve Radius. Overall, the
360
Defensive controller remained the most comfortable and natural, compared to the Aggressive and
361
Turner controllers, indicated by the odds ratios for all Village sections. In contrast, not much
362
difference was seen in the evaluation for comfort and naturalness between the Aggressive and
363
Turner controllers, for the Village sections.
364
25
Table 5 - The GEE model parameter estimates and odds ratios, for comfort and naturalness in rural and village roads with different curvatures
365
Comfort
Naturalness
Environment
Curvature
N
Driving style
Coefficient
SE
Sig
OR
N
Driving styles
Coefficient
SE
Sig
OR
Rural
Straight
216
Defensive vs Turner
1.662
0.326
0.000*
5.270
Defensive vs Turner
1.709
0.347
0.000*
5.522
Defensive vs Aggressive
0.969
0.293
0.001*
2.636
215b
Aggressive vs Turner
1.569
0.365
0.000*
4.801
Aggressive vs Turner
0.693
0.366
0.058
1.999
Defensive vs Aggressive
0.140
0.253
0.580
1.150
300to800
288
Defensive vs Turner
1.926
0.289
0.000*
6.860
Defensive vs Turner
1.316
0.307
0.000*
3.727
Defensive vs Aggressive
1.350
0.290
0.000*
3.856
288
Aggressive vs Turner
1.141
0.342
0.001*
3.129
Aggressive vs Turner
0.576
0.310
0.063
1.779
Defensive vs Aggressive
0.175
0.301
0.561
1.191
200to300
72
Defensive vs Turner
2.812
0.483
0.000*
16.650
Defensive vs Turner
4.067
0.624
0.000*
58.356
Aggressive vs Turner
1.422
0.506
0.005*
4.146
72
Aggressive vs Turner
2.441
0.667
0.000*
11.486
Defensive vs Aggressive
1.390
0.411
0.001*
4.016
Defensive vs Aggressive
1.625
0.453
0.000*
5.081
150to200
72
Defensive vs Turner
2.800
0.496
0.000*
16.446
Aggressive vs Turner
2.728
0.677
0.000*
15.295
Aggressive vs Turner
1.561
0.522
0.003*
4.764
72
Defensive vs Turner
2.670
0.458
0.000*
14.439
Defensive vs Aggressive
1.239
0.488
0.011*
3.452
Defensive vs Aggressive
-0.058
0.547
0.916
0.944
less150
288
Defensive vs Turner
1.667
0.252
0.000*
5.295
Defensive vs Turner
1.784
0.303
0.000*
5.955
Defensive vs Aggressive
0.966
0.277
0.000*
2.627
287c
Aggressive vs Turner
1.159
0.415
0.005*
3.185
Aggressive vs Turner
0.701
0.269
0.009*
2.016
Defensive vs Aggressive
0.626
0.265
0.018*
1.869
Village d
Straight
72
Defensive vs Turner
1.859
0.459
0.000*
6.415
Aggressive vs Turner
0.958
0.410
0.020*
2.607
Defensive vs Aggressive
1.342
0.441
0.002*
3.828
72
Defensive vs Turner
0.083
0.5525
0.880
1.087
Aggressive vs Turner
0.516
0.427
0.227
1.676
Defensive vs Aggressive
-0.875
0.508
0.085
0.417
300to800
72
Defensive vs Aggressive
2.934
0.767
0.000*
18.796
Defensive vs Aggressive
3.667
0.542
0.000*
39.121
Defensive vs Turner
2.480
0.425
0.000*
11.936
72
Defensive vs Turner
1.917
0.513
0.000*
6.798
Aggressive vs Turner
-0.454
0.554
0.412
0.635
Aggressive vs Turner
-1.75
0.800
0.029*
0.174
200to300
71a
Defensive vs Turner
3.591
0.514
0.000*
36.255
Defensive vs Turner
3.917
0.692
0.000*
50.233
Aggressive vs Turner
2.159
0.529
0.000*
8.662
72
Aggressive vs Turner
2.75
0.660
0.000*
15.643
Defensive vs Aggressive
1.432
0.552
0.009*
4.186
Defensive vs Aggressive
1.167
0.631
0.064
3.211
150to200
144
Defensive vs Turner
2.191
0.380
0.000*
8.948
Defensive vs Turner
1.396
0.410
0.001*
4.038
Defensive vs Aggressive
1.649
0.440
0.000*
5.204
144
Defensive vs Aggressive
1.021
0.377
0.007*
2.776
Aggressive vs Turner
0.542
0.372
0.145
1.719
Aggressive vs Turner
0.375
0.537
0.485
1.455
less150
144
Defensive vs Turner
2.962
0.431
0.000*
19.336
Defensive vs Turner
3.042
0.531
0.000*
20.940
Defensive vs Aggressive
2.943
0.529
0.000*
18.974
144
Defensive vs Aggressive
2.438
0.726
0.001*
11.444
Aggressive vs Turner
0.238
0.303
0.432
1.268
Aggressive vs Turner
0.604
0.790
0.444
1.830
Note. * p < .05. Orders of comparison are based on odds ratios. a b c 1 observation was missing. d GEE with the linear link function was used for naturalness (see Appendix 2).
366
26
The influence of personality traits on subjective evaluation
367
Following data collection, participants were divided into two sub-samples, based on their average
368
scores to the 20 AISS items. Evaluation of the controllers by the two sub-samples, providing the
369
lowest (mean = 48.54, N = 13), and highest sensation seeking score (mean = 59.45, N = 11), was then
370
assessed.
371
Table 6 and Figure 9 show that the Defensive driving style was regarded as the most comfortable, for
372
both the high and low sensation seekers. Interesting results were observed regarding the evaluation
373
of naturalness. Low sensation seekers evaluated the Defensive as much more natural than the other
374
two controllers, whereas high sensation seekers rated the Aggressive and Defensive driving styles
375
about the same, in terms of naturalness. This finding also explains the bimodal pattern of evaluations
376
on naturalness, shown in Figure 5 (right).
377
Table 6 - The GEE model parameter estimates and odds ratios regarding comfort and naturalness for
378
low and high sensation seekers
379
Comfort
N
Driving style
Coefficient
SE
Wald
Sig
Odds ratio (OR)
Low
sensation
seekers
934a
Defensive vs Turner
1.809
0.346
27.305
0.000*
6.102
Defensive vs Aggressive
1.356
0.378
12.875
0.000*
3.881
Aggressive vs Turner
0.452
0.436
1.075
0.300
1.572
High
sensation
seekers
791b
Defensive vs Turner
2.153
0.296
53.039
0.000*
8.607
Defensive vs Aggressive
1.507
0.281
28.865
0.000*
4.515
Aggressive vs Turner
0.645
0.348
3.438
0.064
1.906
Naturalness
N
Driving style
Coefficient
SE
Wald
Sig
Odds ratio (OR)
Low
sensation
seekers
936
Defensive vs Turner
1.593
0.262
37.061
0.000*
4.919
Defensive vs Aggressive
0.949
0.368
6.666
0.010*
2.583
Aggressive vs Turner
0.644
0.463
1.933
0.164
1.904
High
sensation
seekers
789c
Defensive vs Turner
1.652
0.379
18.98
0.000*
5.217
Aggressive vs Turner
1.258
0.370
11.584
0.001*
3.518
Defensive vs Aggressive
0.394
0.334
1.390
0.238
1.483
Note. * p < .05. abc observations were missing, with the number of 2, 1, and 3, respectively.
380
27
381
Figure 9. Overall comfort (left) and naturalness (right) evaluation of the driving styles from different
382
sensation seekers.
383
4. Discussion
384
This driving simulator study examined users’ subjective evaluation of the driving style of three AV
385
controllers, in terms of comfort and naturalness, when negotiating a range of rural and village
386
sections of a UK road. The link between participants’ sensation seeking scores and their evaluation
387
of these controllers was also investigated.
388
In terms of human- vs machine-like controllers, results showed that users preferred the two human-
389
like AV controllers, in terms of both comfort and naturalness. Contrasting our findings with similar
390
studies is challenging, as, at the time of writing, there are very few studies which have explicitly
391
compared participant preferences for human-like and machine-like automated vehicle controllers.
392
One exception is a study by Oliveira et al. (2019), who measured users’ trust towards a highly
393
28
automated pod, which showed either human- or machine-like driving behaviours, when crossing a T-
394
junction. In this study, human-like behaviour was produced by demonstrating a cautious “peeking”
395
behaviour by the pod, before it crossed the junction, while machine-like behaviour was produced by
396
an assertive crossing, as if the road conditions were known to the automated pod. Oliveira et al.
397
(2019) showed no difference in trust ratings for the two behaviours of the pod. There are two reasons
398
why our study results are in contrast to those of Oliveira et al. (2019). One may be due to a difference
399
in the concept used between our two studies: trust versus comfort and naturalness. The other may
400
be because of the lower operating speed of the AV used in an urban road, by Oliveira et al, compared
401
to the higher travelling speeds of our vehicle, travelling in rural road sections. This contrast in results
402
illustrates the importance of considering the scenarios used to evaluate AV driving styles in such
403
studies, since they vary across different road environments, based on both geometry and posted
404
speed limit, which clearly influences any subjective assessment and evaluation (Hajiseyedjavadi et al.,
405
submitted). Further work on the influence of different scenarios on subjective appraisals of human-
406
vs machine-like AV driving styles, should clarify this.
407
Overall, participants rated the Defensive controller more comfortable than the other two controllers,
408
while both the Defensive and Aggressive controllers were assessed as more natural than the Turner.
409
This suggests that there may be a distinction between what human evaluators consider a comfortable
410
versus natural driving style, which is perhaps in contrast to the suggestion made by Elbanhawi et al.
411
(2015), who regarded natural, or familiar, driving manoeuvres as one contributor to driving comfort.
412
Our results suggest that comfort and naturalness of a controller should not be used interchangeably
413
in such research, and that while human-like driving styles can be considered as equally more natural
414
than a ML-based controller, they are not necessarily as equally comfortable. Therefore, factors which
415
29
contribute to the comfort of a controller are not the same as those that determine its naturalness.
416
Regarding how road geometries and vehicle kinematics affected subjective ratings, our results show
417
that variations in speed potentially had a greater influence on evaluation of comfort and naturalness
418
of the controllers, when compared to differences in lateral offset. This was especially the case for the
419
rural sections, which contained roads of tighter curvature, and higher speed. These results are in line
420
with the work of Hajiseyedjavadi et al. (submitted), who found that their model-based human-like
421
AV driving controllers were assessed as less pleasant when negotiating narrower curves. These
422
authors also found that a more rigid controller, which always followed the centre of the lane, received
423
better evaluations. Together, these results suggest the influence of vehicle kinematics and road
424
geometry on subjective evaluation of AV controllers. Moreover, our results showed little difference
425
in lateral kinematic features of the three controllers, which was also reflected in the evaluations.
426
Therefore, future studies need to examine the effect of more pronounced lateral offset on subjective
427
evaluation, especially since maintaining sufficient and safe distance to road edges is thought to
428
enhance driving comfort (Summala, 2007).
429
We found an interesting interaction between personality trait and evaluation of the controllers, with
430
high sensation seekers rating the Aggressive driving style (which was a recording of another
431
representative high sensation seeker) as natural, which was not the case for low sensation seekers.
432
As naturalness in this study was defined as a driving style that is closest to your own driving”, it is
433
interesting to see this strong influence of personality traits on driving style and preference. The
434
distinction between comfort and naturalness as concepts is also highlighted here, because there was
435
no difference in the two groups, when evaluating the comfort of the Aggressive driving style. In other
436
words, while the high sensation seekers thought the Aggressive driving style was natural, they did
437
30
not find it comfortable. These results highlight the value of personalisation of automated controllers,
438
to benefit the range of preferences by consumers with varying personality traits, notwithstanding
439
their safety considerations.
440
Limitations
441
One limitation of the present study is the motion-planning performance of the Turner controller,
442
which was developed using a small number of participants. Moreover, although the motion planner’s
443
output consists of a series of aim speeds and positions, we only used simple controllers that were
444
manually calibrated for the automated vehicle to drive, using this data flow. Thus, a future study
445
could use more data to train the motion planner and consider a better approach for implementing
446
the controllers.
447
As with all controlled driving simulator studies, there are caveats regarding the relevance and
448
generalisability of these findings, and their implications with respect to real-world AV controllers.
449
Creating very realistic controllers was possible in this driving simulator study, due to its advanced
450
motion-controller capabilities. However, future studies would benefit from evaluating these sorts of
451
controllers in more real-world settings, also assessing how such evaluation is affected by other real-
452
world factors, such as different road surfaces, or presence of other roadside objects and road users.
453
Conclusions
454
Participants rated the two human-like driving styles as more natural, compared with the less human-
455
like, ML-based, controller. Most participants also rated the Defensive driving style (gentler speed
456
profiles) as more comfortable than the Aggressive controller (higher accelerations and more sudden
457
31
braking profiles). This study shows, for the first time, that participants are able to distinguish between
458
the natural driving manoeuvres of humans and the more machine-like negotiations of an artificial
459
controller. In addition, we illustrate that there is a more complex relationship between concepts such
460
as comfort and naturalness when evaluating automated vehicle controllers.
461
Acknowledgements
462
The first author is funded by the European Union Horizon 2020 funded SHAPE-IT project (Grant
463
number: 860410). This study was part of the HumanDrive project, funded by Innovate UK and the
464
Centre for Connected and Automated Vehicles (Project number TS/P012035/1). We would like to
465
thank Andrew Tomlinson, Anthony Horrobin and Michael Daly for their implementation of the
466
simulator scenarios.
467
Key points
468
Human-like driving features were rated as more comfortable and natural than the ML-based,
469
less-human-like, driving style, under most road contexts for users with different sensation
470
seeking traits.
471
Between the two human-like driving controllers, the slower, milder and more conservative
472
controller (Defensive) was preferred, in terms of comfort and naturalness.
473
When exposed to more complicated road environments (e.g., sharper curves, higher speed
474
limits, potentially hazardous roadside objects), the Defensive driving controller was preferred.
475
A natural driving style was not necessarily perceived as comfortable, especially by high
476
sensation seekers.
477
32
References
478
Arnett, J. (1994). Sensation seeking: A new conceptualization and a new scale. Personality and
479
Individual Differences, 16(2), 289296.
480
Beggiato, M., Hartwich, F., & Krems, J. (2019). Physiological correlates of discomfort in automated
481
driving. Transportation Research Part F: Traffic Psychology and Behaviour, 66, 445458.
482
https://doi.org/10.1016/j.trf.2019.09.018
483
Bellem, H., Thiel, B., Schrauf, M., & Krems, J. F. (2018). Comfort in automated driving: An analysis of
484
preferences for different automated driving styles and their dependence on personality traits.
485
Transportation Research Part F: Traffic Psychology and Behaviour, 55, 90100.
486
https://doi.org/10.1016/j.trf.2018.02.036
487
Butakov, V., & Ioannou, P. (2015). Driving Autopilot with Personalization Feature for Improved Safety
488
and Comfort. IEEE Conference on Intelligent Transportation Systems, Proceedings, ITSC, 2015-
489
Octob, 387393. https://doi.org/10.1109/ITSC.2015.72
490
Carsten, O., & Martens, M. H. (2018). How can humans understand their automated cars? HMI
491
principles, problems and solutions. Cognition, Technology and Work, 21(1), 320.
492
https://doi.org/10.1007/s10111-018-0484-0
493
Ebe, K., & Griffin, M. J. (2001). Factors affecting static seat cushion comfort. Ergonomics, 44(10), 901
494
921. https://doi.org/10.1080/00140130110064685
495
Elander, J., West, R., & French, D. (1993). Behavioral Correlates of Individual Differences in Road-
496
Traffic Crash Risk: An Examination of Methods and Findings. Psychological Bulletin, 113(2), 279
497
294. https://doi.org/10.1037/0033-2909.113.2.279
498
Elbanhawi, M., Simic, M., & Jazar, R. (2015). In the Passenger Seat: Investigating Ride Comfort
499
Measures in Autonomous Cars. IEEE Intelligent Transportation Systems Magazine, 7(3), 417.
500
https://doi.org/10.1109/MITS.2015.2405571
501
Ge, Y., Qu, W., Jiang, C., Du, F., Sun, X., & Zhang, K. (2014). The effect of stress and personality on
502
dangerous driving behavior among Chinese drivers. Accident Analysis and Prevention, 73, 34
503
40. https://doi.org/10.1016/j.aap.2014.07.024
504
Hajiseyedjavadi, F., Romano, R., Paschalidis, E., Wei, C., Solernou, A., Jamson, A. H., Boer, E. R., &
505
Merat, N. (2021). Effect of Environmental Factors and Individual Differences on Subjective
506
Experience of Human-Like and Conventional Automated Vehicle Controllers [Manuscript
507
submitted for publication]. 44(0), 122. https://doi.org/10.13140/RG.2.2.13778.68808
508
Hartwich, F., Beggiato, M., & Krems, J. F. (2018). Driving comfort, enjoyment and acceptance of
509
automated driving–effects of drivers’ age and driving style familiarity. Ergonomics, 61(8), 1017
510
1032. https://doi.org/10.1080/00140139.2018.1441448
511
33
Jamson;, A. H., Horrobin, A. J. ., & Auckland, R. A. (2007). Whatever Happened to the LADS? Design
512
and development of the new University of Leeds Driving Simulator. DSC 2007 North America.
513
Kaber, D. B., & Endsley, M. R. (2004). The effects of level of automation and adaptive automation on
514
human performance, situation awareness and workload in a dynamic control task. In Theoretical
515
Issues in Ergonomics Science (Vol. 5, Issue 2). https://doi.org/10.1080/1463922021000054335
516
Liang, K. Y., & Zeger, S. L. (1986). Longitudinal data analysis using generalized linear models.
517
Biometrika, 73(1), 1322. https://doi.org/10.1093/biomet/73.1.13
518
Louw, T., Hajiseyedjavadi, F. H., Jamson, H., Romano, R., Boer, E., & Merat, N. (2019). The Relationship
519
between Sensation Seeking and Speed Choice in Road Environments with Different Levels of Risk.
520
April, 2935. https://doi.org/10.17077/drivingassessment.1671
521
Madigan, R., Louw, T., Dziennus, M., Graindorge, T., Ortega, E., Graindorge, M., & Merat, N. (2016).
522
Acceptance of Automated Road Transport Systems (ARTS): An Adaptation of the UTAUT Model.
523
Transportation Research Procedia, 14(0), 22172226.
524
https://doi.org/10.1016/j.trpro.2016.05.237
525
Menner, M., Berntorp, K., Zeilinger, M. N., & Di Cairano, S. (2019). Inverse Learning for Human-
526
Adaptive Motion Planning. Proceedings of the IEEE Conference on Decision and Control, 2019-
527
Decem, 809815. https://doi.org/10.1109/CDC40024.2019.9030020
528
Murphey, Y. L., Milton, R., & Kiliaris, L. (2009). Driver’s style classification using jerk analysis. 2009
529
IEEE Workshop on Computational Intelligence in Vehicles and Vehicular Systems, CIVVS 2009 -
530
Proceedings, 2328. https://doi.org/10.1109/CIVVS.2009.4938719
531
Oliveira, L., Proctor, K., Burns, C. G., & Birrell, S. (2019). Driving style: How should an automated
532
vehicle behave? Information (Switzerland), 10(6), 120. https://doi.org/10.3390/INFO10060219
533
Oppenheim, I., Oron-Gilad, T., Parmet, Y., & Shinar, D. (2016). Can traffic violations be traced to
534
gender-role, sensation seeking, demographics and driving exposure? Transportation Research
535
Part F: Traffic Psychology and Behaviour, 43, 387395.
536
https://doi.org/10.1016/j.trf.2016.06.027
537
Paddeu, D., Parkhurst, G., & Shergold, I. (2020). Passenger comfort and trust on first-time use of a
538
shared autonomous shuttle vehicle. Transportation Research Part C, 115(March), 102604.
539
https://doi.org/10.1016/j.trc.2020.02.026
540
Pennig, S., Quehl, J., & Rolny, V. (2012). Effects of aircraft cabin noise on passenger comfort.
541
Ergonomics, 55(10), 12521265. https://doi.org/10.1080/00140139.2012.703698
542
Price, M., Venkatraman, V., Gibson, M., Lee, J., & Mutlu, B. (2016). Psychophysics of Trust in Vehicle
543
Control Algorithms. SAE Technical Papers. https://doi.org/10.4271/2016-01-0144
544
Qatu, M. S. (2012). Recent research on vehicle noise and vibration. International Journal of Vehicle
545
Noise and Vibration, 8(4), 289301. https://doi.org/10.1504/IJVNV.2012.051536
546
34
Radhakrishnan, V., Merat, N., Louw, T., Lenné, M. G., Romano, R., Paschalidis, E., Hajiseyedjavadi, F.,
547
Wei, C., & Boer, E. R. (2020). Measuring drivers’ physiological response to different vehicle
548
controllers in highly automated driving (HAD): Opportunities for establishing real-time values of
549
driver discomfort. Information (Switzerland), 11(8), 114.
550
https://doi.org/10.3390/INFO11080390
551
Ramm, S., Giacomin, J., Robertson, D., & Malizia, A. (2014). A first approach to understanding and
552
measuring naturalness in driver-car interaction. AutomotiveUI 2014 - 6th International
553
Conference on Automotive User Interfaces and Interactive Vehicular Applications, in Cooperation
554
with ACM SIGCHI - Proceedings, c. https://doi.org/10.1145/2667317.2667416
555
Rehder, E., Quehl, J., & Stiller, C. (2017). Driving Like a Human: Imitation Learning for Path Planning
556
using Convolutional Neural Networks. International Conference on Robotics and Automation
557
Workshops. http://ppniv17.irccyn.ec-nantes.fr/session1/Rehder/paper.pdf
558
Riendeau, J., Stinchcombe, A., Weaver, B., & Bédard, M. (2018). Personality factors are associated
559
with simulated driving outcomes across the driving lifespan. Transportation Research Part F:
560
Traffic Psychology and Behaviour, 54, 171187. https://doi.org/10.1016/j.trf.2018.01.022
561
Rossner, P., & Bullinger, A. C. (2020). How Do You Want to be Driven? Investigation of Different
562
Highly-Automated Driving Styles on a Highway Scenario. Advances in Intelligent Systems and
563
Computing, 964, 3643. https://doi.org/10.1007/978-3-030-20503-4_4
564
SAE International. (2016). Taxonomy and Definitions for Terms Related to Driving Automation
565
Systems for On-Road Motor Vehicles. SAE International.
566
https://doi.org/https://doi.org/10.4271/J3016_201806
567
Siebert, F. W., Oehl, M., Höger, R., & Pfister, H. R. (2013). Discomfort in Automated Driving - The
568
Disco-Scale. Communications in Computer and Information Science, 374(PART II), 337341.
569
https://doi.org/10.1007/978-3-642-39476-8_69
570
Solernou, A., Romano, R., Souflas, I., Hajiseyedjavadi, F., & Paschalidis, E. (2020). Integration and
571
training of a ROS autonomous driver for human-like driving style in a complex multi-component
572
driving simulator. DSC 2020 Europe, 911.
573
Summala, H. (2007). Towards Understanding Motivational and Emotional Factors in Driver Behaviour:
574
Comfort Through Satisficin. In Modelling Driver Behaviour in Automotive Environments: Critical
575
Issues in Driver Interactions with Intelligent Transport Systems (pp. 189207).
576
https://doi.org/10.1007/978-1-84628-618-6
577
Taubman-Ben-Ari, O., Mikulincer, M., & Gillath, O. (2004). The multidimensional driving style
578
inventory - Scale construct and validation. Accident Analysis and Prevention, 36(3), 323332.
579
https://doi.org/10.1016/S0001-4575(03)00010-1
580
Ulleberg, P., & Rundmo, T. (2003). Personality, attitudes and risk perception as predictors of risky
581
driving behaviour among young drivers. Safety Science, 41(5), 427443.
582
https://doi.org/10.1016/S0925-7535(01)00077-7
583
35
Urmson, C., Anhalt, J., Bagnell, D., Baker, C., Bittner, R., Clark, M. N., Dolan, J., Duggins, D., Galatali,
584
T., Geyer, C., Gittleman, M., Harbaugh, S., Hebert, M., Howard, T. M., Kolski, S., Kelly, A.,
585
Likhachev, M., McNaughton, M., Miller, N., … Ziglar, J. (2008). A Comparison of Automatic Nap-
586
of-the-earth Guidance Strategies for Helicopters. Journal of Field Robotics, 25(8), 425466.
587
https://doi.org/10.1002/rob
588
Wei, C., Romano, R., Merat, N., Wang, Y., Hu, C., Taghavifar, H., Hajiseyedjavadi, F., & Boer, E. R.
589
(2019). Risk-based autonomous vehicle motion control with considering human driver’s
590
behaviour. Transportation Research Part C: Emerging Technologies, 107(August), 114.
591
https://doi.org/10.1016/j.trc.2019.08.003
592
Yusof, N. M., Karjanto, J., Terken, J., Delbressine, F., Hassan, M. Z., & Rauterberg, M. (2016). The
593
exploration of autonomous vehicle driving styles: Preferred longitudinal, lateral, and vertical
594
accelerations. AutomotiveUI 2016 - 8th International Conference on Automotive User Interfaces
595
and Interactive Vehicular Applications, Proceedings, 245252.
596
https://doi.org/10.1145/3003715.3005455
597
Zuckerman, M., & Neeb, M. (1980). Demographic influences in sensation seeking and expressions of
598
sensation seeking in religion, smoking and driving habits. Personality and Individual Differences,
599
1(3), 197206. https://doi.org/https://doi.org/10.1016/0191-8869(80)90051-3
600
Biographies
601
Chen Peng is a PhD candidate and Marie Curie Early Stage Researcher at the Institute for
602
Transport Studies, University of Leeds. She received her Master in Human Technology Interaction
603
from the Eindhoven University of Technology in 2019.
604
Natasha Merat is the research group leader and Chair Professor of the Human Factors and
605
Safety Group at the Institute for Transport Studies, University of Leeds. She received her PhD in
606
Psychology from the University of Leeds in 1998.
607
Richard Romano is the Chair Professor of Driving Simulation at the Institute for Transport
608
Studies, University of Leeds. He received his PhD in Industrial Engineering from the University of Iowa
609
in 1999.
610
Foroogh Hajiseyedjavadi is a Research Fellow at the Institute for Transport Studies, University
611
36
of Leeds. She received her PhD in Transportation Engineering from the University of Massachusetts
612
Amherst in 2018.
613
Evangelos Paschalidis is a Research Fellow at the Institute for Transport Studies, University of
614
Leeds. He received his PhD in Transport Studies from the University of Leeds in 2019.
615
Chongfeng Wei is an Assistant Professor at Queen's University Belfast. He received his PhD in
616
Mechanical Engineering from University of Birmingham in 2015.
617
Vishnu Radhakrishnan is a PhD candidate at the Institute for Transport Studies, University of
618
Leeds. He received his Master in Automotive Engineering from University of Leeds in 2017.
619
Albert Solernou is the Virtuocity Suite Facilities Manager and Simulator Software Architect at
620
the Institute for Transport Studies, University of Leeds. He received his PhD in Chemistry from the
621
Barcelona Supercomputing Centre in 2011.
622
Deborah Forster is a primatologist and cognitive scientist. She received her PhD in Cognitive
623
Science from the University of California San Diego in 2012.
624
Erwin Boer is the President of Entropy Control, Inc. and an Associate Professor in Delft
625
University of Technology. He received his PhD in Electrical and Electronics Engineering from the
626
University of Illinois at Chicago in 1995.
627
37
Appendices
628
Appendix 1 - Cluster analysis for categorising driving behaviour
629
From a previous study of the project, driving behaviours were collected from a sharp curve (radius <
630
150m), a zone with parked cars (length = 162.68m), and the entire drive. The variables used as
631
behavioural indicators for clustering were: root mean square of speed, standard deviation of
632
longitudinal acceleration, and standard deviation of yaw rate. The k-means clustering analysis was
633
conducted.
634
Appendix 2 - Configuration of the GEE model
635
The working correlation matrix was specified as exchangeable, which characterises the correlation
636
structure of multiple observations within a participant as the same. As the comfort and naturalness
637
were rated using an ordinal Eleven-point Likert scale, the distribution of the dependent variable was
638
specified as a multinomial distribution. A link function is used to characterise the relationship
639
between the mean of the response (i.e., subjective ratings) and the linear predictor (i.e., controllers).
640
The ordered logit regression was specified as the link function.
641
However, in some Village road segments, including the Straight road sections, the 200-300 Radius
642
sections, and the 150-200 Radius sections, the statistical model did not provide valid results, because
643
the participants’ responses to the three controllers in these sections showed very similar patterns,
644
which resulted in collinearity. For example, in the Curve Radius 150-200 section, most responses
645
clustered between 2 to 5 for all three driving styles. Therefore, we treated the data as continuous in
646
these road sections, to allow statistical comparisons. The distribution was specified as normal in the
647
38
GEE.
648
Appendix 3 - Vehicle kinematics inspection for the two village road sections with unexpected
649
assessments
650
Inspection of the vehicle-based metrics for these two particular road sections (Figure 10) showed
651
that the Aggressive controller’s speed was markedly higher than that of the two other controllers in
652
the 300-800 Radius section, and higher than the designated speed limit of 30 mph. A sudden
653
fluctuation of speed for the Turner controller was also seen in the 200-300 Radius section, although
654
it was within the speed limit of 40 mph. Further inspection of the simulated scene did not show the
655
presence of any unusually placed road furniture, such as parked cars. A possible explanation here is
656
that the Turner controller did not look far enough ahead to smooth out the speed changes, and was
657
also inadequately sensitive to roadside furniture. Regardless, these results show that the effect of
658
these kinematic changes were clearly felt by our users, which can possibly explain their evaluation of
659
the controllers for these two sections.
660
661
Figure 10. Vehicle speed of three controllers in the two village road curves
662
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
This study investigated how driver discomfort was influenced by different types of automated vehicle (AV) controllers, compared to manual driving, and whether this response changed in different road environments, using heart-rate variability (HRV) and electrodermal activity (EDA). A total of 24 drivers were subjected to manual driving and four AV controllers: two modelled to depict “human-like” driving behaviour, one conventional lane-keeping assist controller, and a replay of their own manual drive. Each drive lasted for ~15 min and consisted of rural and urban environments, which differed in terms of average speed, road geometry and road-based furniture. Drivers showed higher skin conductance response (SCR) and lower HRV during manual driving, compared to the automated drives. There were no significant differences in discomfort between the AV controllers. SCRs and subjective discomfort ratings showed significantly higher discomfort in the faster rural environments, when compared to the urban environments. Our results suggest that SCR values are more sensitive than HRV-based measures to continuously evolving situations that induce discomfort. Further research may be warranted in investigating the value of this metric in assessing real-time driver discomfort levels, which may help improve acceptance of AV controllers.
Article
Full-text available
The selected motions of autonomous vehicles (AVs) are subject to the constraints from the surrounding traffic environment, infrastructure and the vehicle’s dynamic capabilities. Normally, the motion control of the vehicle is composed of trajectory planning and trajectory following according to the surrounding risk factors, the vehicles’ capabilities as well as tyre/road interaction situations. However, pure trajectory following with a unique path may make the motion control of the vehicle be too careful and cautious with a large amount of steering effort. To follow a planned trajectory, the AVs with the traditional path-following control algorithms will correct their states even if the vehicles have only a slight deviation from the desired path or the vehicle detects static infrastructure like roadside trees. In this case, the safety of the AVs can be guaranteed to some degree, but the comfort and sense of hazards for the drivers are ignored, and sometimes the AVs have unusual motion behaviours which may not be acceptable to other road users. To solve this problem, this study aims to develop a safety corridor-based vehicle motion control approach by investigating human-driven vehicle behaviour and the vehicle’s dynamic capabilities. The safety corridor is derived by the manoeuvring action feedback of actual drivers as collected in a driving simulator when presented with surrounding risk elements and enables the AVs to have safe trajectories within it. A corridor-based Nonlinear Model Predictive Control (NMPC) has been developed which controls the vehicle state to achieve a smooth and comfortable trajectory while applying trajectory constraints using the safety corridor. The safety corridor and motion controller are assessed using four typical scenarios to show that the vehicle has a human-like or human-oriented behaviour which is expected to be more acceptable for both drivers and other road users.
Conference Paper
Full-text available
This paper presents the results of a driving simulator study conducted for the UK-funded HumanDrive project, which aims to develop natural, human-like autonomous vehicle control. As part of that effort, this paper examines whether the established relationship between different sensation seeking (SS) traits and speed choice holds true across a range of driving scenarios, with different levels of contextual risk. Risk was introduced by varying a number of factors, including the environment (rural/urban), and the road edge context (low risk, static risk, potentially dynamic risk). Correlation analysis was performed between sensation seeking and the 95th percentile of vehicle speed for roads with different levels of risk, also considering age and gender. The results indicated that, overall, SS was significantly positively correlated with the 95th percentile of vehicle speed, and particularly for drivers under 40 years. SS was also found to correlate positively with speed choice at all risk levels, however, the effect was more pronounced in road environments that were classified as less risky. These findings have design implications for the development of autonomous vehicle control models.
Article
Full-text available
As long as vehicles do not provide full automation, the design and function of the Human Machine Interface (HMI) is crucial for ensuring that the human “driver” and the vehicle-based automated systems collaborate in a safe manner. When the driver is decoupled from active control, the design of the HMI becomes even more critical. Without mutual understanding, the two agents (human and vehicle) will fail to accurately comprehend each other’s intentions and actions. This paper proposes a set of design principles for in-vehicle HMI and reviews some current HMI designs in the light of those principles. We argue that in many respects, the current designs fall short of best practice and have the potential to confuse the driver. This can lead to a mismatch between the operation of the automation in the light of the current external situation and the driver’s awareness of how well the automation is currently handling that situation. A model to illustrate how the various principles are interrelated is proposed. Finally, recommendations are made on how, building on each principle, HMI design solutions can be adopted to address these challenges.
Article
Autonomous Vehicles (AV) may become widely diffused as a road transport technology around the world. However, two conditions of successful adoption of AVs are that they must be synchronously shared, to avoid negative transport network and environmental consequences, and that high levels of public acceptance of the technology must exist. The implications of these two conditions are that travellers must accept sharing rides with unfamiliar others in Shared Autonomous Vehicles (SAV). Two factors that have been identified as being positive influencers of acceptance are comfort and trust. The present paper undertakes a novel examination as to how comfort and trust ratings are affected by specific attributes of the ride experience of travelling in a fully-automated real-world, shared vehicle. To this end, 55 participants experienced riding in an SAV shuttle under experimental conditions at a test facility. Each experimental run involved two unrelated participants, accompanied by a safety operative and a researcher, undertaking four trips in the SAV, during which two conditions were presented for each of the independent variables of ‘direction of face’ (forwards/backwards) and ‘maximum vehicle speed’ (8/16 km/h). Order of presentation was varied between pairs of participants. After each run, participants rated the dependent variables ‘trust’ and ‘comfort’ (the latter variable comprised by six comfort factors). Expected and evaluative ratings were also obtained during pre-experimental orientation and debriefing sessions. Statistically significant relationships (p < .001) were found between trust and each of the independent variables, but for neither variable in the case of perceived comfort. A strong correlation was found between comfort and trust, interpreted as indicating trust in the SAV as an important predictor of perceived comfort. The before and after-experiment ratings for both variables showed statistically significant increases, and particularly for daily car drivers.
Article
Comfort in automated driving is considered a key issue for broad acceptance of automated vehicles. The research project KomfoPilot at Chemnitz University of Technology aimed to investigate factors that influence comfort in automated driving as well as to identify physiological indicators of discomfort. In an empirical driving simulator study, 40 participants from 25 to 84 years old experienced three highly automated trips including six potentially uncomfortable situations in each trip. Participants reported perceived discomfort continuously by a handset control. The physiological parameters Heart Rate (HR) and Electrodermal Activity (EDA) were assessed using the smartband Microsoft Band 2; pupil diameter and eye blinks were measured by the SMI Eye Tracking Glasses 2. Results showed specific physiological reactions in situations that provoked moderate to high discomfort. Longer lasting and slowly evolving situations with lower reported discomfort did not show associated changes in physiological parameters. HR decreased consistently during uncomfortable situations, which could be related to the phenomenon “preparation for action”. Pupil diameter increased and eye blink rate decreased in uncomfortable situations that were visually monitored. EDA did not show specific effects, which, however could be attributed to measurement procedures of the smartband. The results serve as a basis for developing a real-time discomfort detection algorithm and will additionally be validated on-road.
Article
Research has shown that personality factors are related to driving safety. However, the majority of existing studies rely on self-report measures of driving behaviour and sample drivers from limited age ranges. This study sought to examine the relationship between personality and objective driving outcomes as assessed by a driving simulator in a sample of young, mid-aged, and older adults. A total of 114 active drivers completed personality questionnaires as well as a simulated driving assessment protocol. The results showed that: (1) Extraversion and neuroticism were significantly associated with driving simulator performance; (2) conscientiousness was significantly associated with driving performance among middle-aged adults; (3) sensation seeking was an important personality factor primarily for young drivers and was positively correlated with driving speed in the simulator. These results provide further support for the link between personality factors and driving performance, and suggest certain directions for future research.
Article
Automated driving has the potential to improve the safety and efficiency of future traffic and to extend elderly peoples’ driving life, provided it is perceived as comfortable and joyful and is accepted by drivers. Driving comfort could be enhanced by familiar automated driving styles based on drivers’ manual driving styles. In a two-stage driving simulator study, effects of driving automation and driving style familiarity on driving comfort, enjoyment, and system acceptance were examined. Twenty younger and twenty older drivers performed a manual and four automated drives of different driving style familiarity. Acceptance, comfort, and enjoyment were assessed after driving with standardised questionnaires, discomfort during driving via handset control. Automation increased both age groups’ comfort, but decreased younger drivers’ enjoyment. Younger drivers showed higher comfort, enjoyment and acceptance with familiar automated driving styles, whereas older drivers preferred unfamiliar, automated driving styles tending to be faster than their age-affected manual driving styles. Practitioner Summary Automated driving needs to be comfortable and enjoyable to be accepted by drivers, which could be enhanced by driving style individualisation. This approach was evaluated in a two-stage driving simulator study for different age groups. Younger drivers preferred familiar driving styles, whereas older drivers preferred driving styles unaffected by age.