Impacts of Different Driving Automation Levels on Highway Geometric Design from the Perspective of Trucks

Truck automation is emerging as an innovative technology with beneﬁts in traﬃc safety and the economy to revolutionize freight traﬃc. Despite these beneﬁts, the potential negative or positive eﬀects of diﬀerent driving automation levels (from no automation to full automation) on highway geometry remained to be determined. In this study, diﬀerences related to sight distance characteristics among varied automation levels were ﬁrstly discussed and calibrated. Then, seven analysis scenarios of typical levels were proposed. Based on each level with tailored characteristics, the current models of geometric design elements including the required stopping sight distance, horizontal sight line oﬀset, and lengths of vertical curves were revised. Finally, impacts of each level on computed values of those elements were evaluated. Results show that high or full driving automation could substantially lower the requirements of geometric design. Active safety systems have a similar role but with less signiﬁcant eﬀects. Diﬀerently, the driver assistance and partial or conditional automation systems put a higher demand on the road geometric design in terms of driving safety. Outcomes of this study can be used to design real-world geometry of dedicated lanes and provide a methodological basis for the operation of diﬀerent driving automation features.


Introduction
Over the last decades, rapid innovations of artificial intelligence and sensor technology within the automotive sector have brought about dramatic advances in the field of transportation. Vehicles equipped with active safety or driving automation systems (AVs) have received substantial attention for their potential positive effects on safety, traffic flow, energy-saving, and driver workload [1][2][3][4]. Besides, the passenger car, truck with automation (AT) is also the primary representative of AV [5].
Due to the considerable reduction in the cost of labor forces [6] and fuels in freight [7], truck automation is expected to penetrate into the market quicker and is more necessary than automation for passenger cars. Also, dedicated truck lanes on highways proposed by Bucklew [8] can serve as the cradle for ATs. However, the potential influence of ATs on highway geometry seems to have received only limited attention to the best of our knowledge [9][10][11][12]. Yet, road transportation, an essential mode of transport, accounts for the largest share of transportation worldwide, and the total length of highways in China reached 4,696,300 km by the end of 2016 [13]. Even a minor impact on highway geometry can therefore have substantial economic and safety implications. Given the rapid growth of driving automation technology, there is a pressing need to uncover its impact on highway geometry. e Society of Automotive Engineers (SAE) International [14] has provided a taxonomy with detailed definitions of six driving automation levels (Level 0 ∼ Level 5, referred to as L0 ∼ L5 hereafter) from none to full driving automation. ese levels can be defined by reference to how strong the automation system intervenes in the dynamic driving tasks (DDT) and/or DDT fallback. Drivers are still responsible for the driving tasks before the system evolves into high or full automation. Although many research projects have alluded to a revolutionary introduction of fully automated driving, the development of driving automation for now is better described as an evolutionary process [15][16][17]. Previous research has demonstrated that automation can take the human driver out of the loop, resulting in deteriorated reactions and performances in the worst-case scenario [18][19][20], for example, system failures or beyond the operational design domain (ODD).
Consideration of driver performance is widely known to be essential to the proper geometric design of highways. e driving behavior of ATs obviously differs from that of human-driven trucks in response time and environmental sensing mode and thus will certainly affect geometry differently [9,10]. To the best of our knowledge, omas and Martinez-Perez [11] were among the first to examine the impact of automated vehicles on highway geometry. ey designed a series of road-train scenarios based on deployments of the well-known Californian Partners for Advanced Transportation and Technology (PATH) and European Safe Road-trains for the Environment (SARTRE). ey carried out a quantitative analysis by assuming a continuum of perception reaction time from 0∼2.5 s with an interval of 0.5 s. Khoury et al. [12] initially investigated the direct effects of a fully autonomous vehicle fleet on geometric design elements including stopping and decision sight distance, and the length of vertical curves. ey all suggested that fully automation systems with the complete elimination of the human driver can achieve significant economic improvements, for e.g., considerable reductions in the length of crest and sag vertical curves. However, little is known about the effects of varied driving automation levels on highway geometric design elements. Given that driving automation systems are becoming more and more important, it is necessary to examine the discrepancy among all automation levels and how this, in turn, affects the driver behavior and highway geometric design, which poses potential safety implications.
Considering that highways in service have saturated available land space, it is uneconomical and impractical to design new geometry for resisting the potential negative effects of AVs [12]. By optimizing sensing solutions or fallback strategies, the negative effects of AVs may decrease if they can detect obstacles in an instant and provide reliable support, for example, instant braking or steering for drivers who cannot take-over promptly. From a technical perspective, when equipped with high-definition maps [21,22], high-tech perception sensors [23], and reliable planning and control algorithms [24], AVs are capable of automatically dealing with risk circumstances and fully adapting to the current practices of highway geometry.
As it remains unknown when and whether fully automated vehicles will come and how the transitional automation levels will affect the geometric design elements, the objective of this study was to quantify the potential negative or positive effects of segregated driving automation levels on highway geometric design from the perspective of trucks. To this end, consideration was given to distinctions in characteristics in terms of sight distances among varied automation levels.
e novel contribution for this study is not the computation model of highway geometric elements itself but rather the methods of including features of all automation levels into the models and new insights into potential impacts from varied automation levels. e remainder of this paper is organized as follows. First, differences related to characteristics of sight distance among varied automation levels are discussed. en, the influence of each automation level with tailored features on highway geometric design elements is calculated and analyzed.
e final section provides conclusions of the work.

Sight-Distance-Related Difference Analysis
First, L0 defined by SAE International [14], is adjusted for our analysis.
en, anticipated performances of human drivers in different automation levels were investigated by describing their possible involved tasks. Based on that, features with respect to the sight distance among varied driving automation levels were elucidated and calibrated by the scrutiny of existing literature and applications.

Revised Level 0 (No Driving Automation).
Design vehicles, including the passenger car, truck, and bus, in many current highway geometric specifications [25,26] do not take any driving automation systems or active safety systems into account. According to [14], such human-driven conventional vehicles with no driving automation (HVs) are considered as L0. Besides, active safety systems, such as forward collision warning (FCW), that merely provide momentary assistance or intervention to help avoid or mitigate potential collisions instead of performing DDT on a sustained basis are also classified as L0. Numerous studies have been conducted to substantiate that warning alerts can promote quicker brake reaction time [27][28][29]. For stopping sight distance analysis, performance (e.g., the reaction/response time to an unexpected event) of drivers in the loop is particularly important. We excluded some kinds of active safety systems including active control of the vehicle subsystems (brake, throttle, suspension, etc.), for example, automated emergency braking (AEB), because such systems are primarily focused on improving safety rather than comfort and convenience that are two of important factors in highway geometric design. erefore, we divided L0 into two sublevels: L0 − and L0 + to denote HVs equipped with neither driving automation nor active safety systems, and HVs only supported by active safety systems, respectively, for difference analysis below.

Driver Tasks and
States from L0 to L5. Figure 1 displays all possible tasks of the driver in L0∼L5 automation condition (driver's right side) and general measurements of his/ her state (driver's left side). Workload and situation awareness are regarded as two of the most important human factors that are predictive of performance and safety [30]. It is known that the driving automation level is determined by an allocation of roles in DDT and its fallback performance between the human driver and driving automation systems. According to listed tasks in Figure 1 and role definitions in [14], when driving automation systems (if any) are engaged, drivers: (1) In L0 − entail no change with those in the conventional vehicle without any automation or active safety systems, i.e., performs task 1∼4 except task 3.3 (2) In L0 + execute the same tasks as those in L0 − but with the support of active safety systems in task 3 (3) In L1 and L2 conduct either task 1 or 2 and neither task 1 nor 2, respectively, and both complete task 3 (with the assistance of driving automation system within its limited dimension, i.e., lateral or/and longitudinal, respectively) and task 4 (4) In L3 are obliged to perform task 4 in a timely manner and are receptive to task 3.2, and are inclined to pick up task 5 (5) In both L4 and L5 need not perform task 1∼4, and are willing to be busy with task 5 as a passenger Reliefs from the cognitive activity (task 3) and the physical activity (task 1 or/and 2) are associated with a decreased workload [31], and distraction from the road is associated with reduced situation awareness [32]. It is evident that the driving automation system at each level (L1∼L5) performs the remainder of tasks (except task 5).
us, drivers tend to engage in fewer tasks with the evolvement of driving automation. However, fewer engaged tasks do not necessarily cause a reduction in workload and situation awareness. Higher automation (L1∼L5) could raise workload and strengthen situation awareness with respect to L0 (L0 − and L0 + ) if the driver has to remain vigilant and monitor/supervise the driving environment (task 3.1) and automation status (task 3.3) [33]. We neglected this divergent result (i.e., higher automation improves the workload and situation awareness) because a wealth of evidence shows that drivers in higher automation condition are more likely to engage in nondriving tasks (task 5) [34] or distraction [35], i.e., higher automation results in lower workload [36] and weaker situation awareness [37]. In addition, results or conclusions of the worst-case scenario at each automation level are the focus in highway geometric study for safety.
It should be noted that although L4 and L5 drivers who can be also considered as passengers perform "poor" in driving tasks, they have already handed over the entire control to automated systems.
at means that drivers' performances do not at all negatively impact vehicle's safe driving controlled by reliable systems. Current practical Automated Driving Systems (ADSs, i.e., L3∼L5) customarily encompass L0 + features. However, what is missing in the literature is the evidence that there is a statistically reliable difference in workload or situation awareness between ADSs with and without the support of L0 + features. Hence, we did not take the possible impact of L0 + on ADSs into consideration. Lastly, both L4 and L5 drivers/ADSs conduct identical tasks when ADSs are engaged. Differences between L4 and L5 only exist in the requirement of the operational design domain that is beyond the scope of this paper. us, we considered these two levels as one (L4/5).

Sight-Distance-Related Characteristics of Actors from
L0∼L5. Actors in driving from L0∼L5 involve the human driver and the driving automation system. eir characteristics with respect to sight distance considered in this paper include perception-brake reaction time, the deceleration rate, the height of the actor's sensing unit (i.e., driver's eye and sensors), and the height of the object.

Perception-Brake Reaction Time.
In order to interpret the perception-brake reaction time (PbRT) of vehicles equipped with L1∼L5 automation, components of PbRT and the stopping process of each level were analyzed. Blocks with the same color possess comparable cognitive or physical properties. For example, sensation is defined as the detection of an object above the roadway, and response selection and programming for humans correspond to planning and decision-making for systems. For L0 − driver, Green [38] decomposed the human PbRT that served as the gauge into a sequence of components (Figure 2(a)): mental processing time (including subcomponents: sensation, perception and response selection, and programming) and movement time. PbRT for computing the required stopping sight distance usually does not add the last device's (i.e., brake pedal) response time [38]. From the description  Figure 1: Schematic view of the driver's tasks in L0∼L5 and measurements of the driver state. Note that (1) nondriving tasks (task 6) do not include those (turning a radio, scanning road signs, answering the phone, etc.) that only require short periods of eyes-off-road time and can be performed simultaneously with driving tasks; (2) strategic tasks (i.e., navigation), such as trip scheduling and route following, were not listed because the difference among automation levels does not include strategic aspects of driving.

External signal occurs
Human driver  above of drivers' tasks in L0 − and L0 + , we derived that the L0 + driver maintains the same PbRT components and structure as those of the L0 − driver. For L1 and L2 vehicles, PbRT can be split into two stages: perception reaction time (PRT) of automation system and PbRT of humans ( Figure 2(b)). Specifically, PRT of the automation system (PRT automation ) denotes perceptual information processing time (including subcomponents: detection, recognition, and classification) of L1 or L2 automation. PbRT of human (PbRT human ) components are the same as those of the L0 driver. It could be possible that PRT automation and PbRT human overlap or initiate simultaneously, i.e., the driver determines the disengagement of the system in advance. e worst-case scenario depicted in Figure 2(b) is that PbRT human occurs after the system disengages. Note that the planning and decision-making block is not included between the sensing and control layers in both L1 and L2 automation control sequence [39] unlike ADS [40] (see Figures 2(c) and 2(d)). us, the system withdraws directly when the sensing outcomes exceed the preset threshold.
For L3 vehicles, PbRTconsists of two phases: PRT of ADS (PRT ADS ) and human take-over time (TOT) (Figure 2(c) adapted from Zeeb et al. [41]). In the generic modular system pipeline of ADS [42], if the system determines to issue the take-over request, it routinely occurs after planning and decision-making. erefore, PRT ADS covers planning and decision-making, and signal transmission time besides perceptual information-processing time. Components of the take-over process during TOT [41] are basically consistent with those of PbRT human but with some revisions. Particularly, sensation also includes the detection of the take-over request apart from the object ahead. e driver may simultaneously return his/her hands to the steering wheel and move his/her feet toward the pedal systems to achieve motor readiness [43]. Although the starting moment criteria of driver intervention are ambiguous [44], such as a 2-degree steering angle or 10% brake pedal position [45], or an absolute steering acceleration of 5 deg/s 2 [46], it is believed that the device response time is included. It should be noted that the L3 system will continue to perform DDT fallback with a preemptive brake until the driver intervenes (if any) or the time budget exhausts.
For ADS-dedicated vehicles (i.e., vehicles operated exclusively by L4/5 ADS), PbRT indicates the time it takes for the L4/5 system to respond to an external signal. As shown in Figures 2(c) and 2(d), both PbRT for L4/5 and PRT ADS for L3 possess similar components. e former, however, is for the decision to brake whilst the latter is for issuing the request.

Deceleration Rate.
e deceleration rate (A d ) elected in measuring braking distance considers approximately 90 percent of all drivers' capabilities to stay within the ego lane and maintain steering control during the braking maneuver on wet surfaces [25]. Explicit in the determination of this A d is the comprehensive assessment of the driving comfort, tirepavement friction level, and capabilities of vehicle braking systems. Notwithstanding the fact that most current vehicle braking systems and friction available on wet pavement surfaces can provide a much larger A d , the threshold of that should be determined by the comfort of drivers or passengers, especially in the truck context. Truck drivers may wish to avoid large accelerations, due to trailers risking jackknifing and fragile goods [47]. For L0∼L3 vehicles, the human driver still accounts for an appreciable role in DDT or DDT fallback. For L4/5 vehicles, the passenger transformed from driver becomes their important "client." us, there is no change in the A d for different levels.

Height of Actor's Sensing Unit.
It is known that sight distance is the distance along a roadway throughout which the object is continuously perceptible to the actor. erefore, the height of the actor's sensing unit and object (see Section 2.3.4) is important for determining sight distance. In L0 condition, the driver's eye is the sensing unit. As described in Section 2.2, the driver still supervises task 3 despite the introduction of L1 or L2 automation. us, the driver's eye is the sensing unit in both L1 and L2 automation as before. For ADSs, it is obvious that perception sensors, such as cameras, Radar, and Lidar [48], become the sensing unit instead. If the driver takes over the control of task 1∼3 ( Figure 1) from ADSs, the human eye becomes the sensing unit again, especially when he/she performs DDT fallback in L3 automation.
For night driving on highways without lighting, the sight distance for L0∼L2 is controlled by the headlight mounting height rather than by the direct line of eye sight. Both Radar and Lidar are independent of light conditions and can work in the darkness [49]. Although many commercial and Journal of Advanced Transportation research vehicles utilize multiple identical sensors and varied sensor combinations (i.e., multisensor fusion) to increase perception reliability and robustness, a broad consensus has been reached about utilizing Lidar in the L4/5 sensor configuration [50]. For L3 automation, whether Lidar should be installed has not been determined due to current high costs or evolved perception technologies. Tesla, for example, is convinced that vision algorithms realized through cameras enable L3 automation sufficiently [50]. Cameras, however, perform poorly at night. Furthermore, the upward divergence angle of the sensing beam from the longitudinal axis of the vehicle should be also tailored to the characteristics of the sensing unit at each level. Particularly, it remains the upward divergence angle of the light beam in the L0∼L2 condition. en, it alters to the upward vertical field of view (FoV) of sensors [51] for ADSs.

Height of Object.
e height of object (h o ) is considered as the representative of the smallest or the shortest object that involves risk to human drivers. Driving automation systems are widely known to be greatly beneficial for preventing accidents caused by human errors [52]. It should be noted that although it is difficult to obtain detailed shape information of distant objects by active sensors (e.g., Lidar and Radar) with the limited vertical resolution [53] (i.e., the obtained point cloud is sparse), recognition using camera images enables to perceive small distant objects successfully. However, to the best of our knowledge, no systematic investigation has studied the possible h o resulting in AV crashes. It is anticipated that the h o for driving automation systems sufficient to provide a safety margin against accidents would increase to a certain extent. Because AVs with dedicated sensors will have an enhanced capability to perceive situations involving collision risk within the perception range, it is anticipated that the h o for driving automation systems sufficient to provide a safety margin against accidents would increase to a certain extent.

Design Criteria for all Automation Levels.
For a general analysis, two much-cited design specifications of highway alignment [25,26] were selected as the basic reference for the design criteria of L0 − .

Perception-Brake Reaction Time.
Although multiple studies have stated that there is no such thing as a single, general PbRT for all drivers [38,41], we adopted the relatively valid fixed PbRT tailored for each automation level. Because sufficient convergence exists among studies of the similar object and methodology to enable reasonable estimates, situation awareness can be inferred from a driver's response to an unexpected event ( Figure 1). As a result, the anticipated trend of PbRT results is in the same direction to that of situation awareness (see Section 2.2). Green's much-cited review [38] declared that the PbRT of 2.50 s adopted in [25] is a reasonable guess for the 90% to 95% general population. Predominantly driven by a trained group of professionals, as opposed to occasional drivers in the car context, truck drivers will response or take back control quicker than passenger car drivers [54]. Because the separate PbRT for trucks and passenger drivers is not generally used in highway design [25], the PbRT of truck drivers in L0 − is still 2.50 s. Bao et al. [55] investigated how a crash warning system influences truck drivers' brake reaction time (i.e., PbRT). Shorter PbRT (mean � 1.62 s) was observed in drivers with alert signals, which correspond to the PbRT (1.64 s) for alerted drivers in [25]. Accordingly, we adopted the PbRT of 1.62 s in L0 + .
Onboard vehicle detection systems have high computational requirements as they need to process the acquired sensing data in real-time for driver/system reactions. For nontrivial velocities of AVs, processing latency should be no larger than 100 ms typically [56]. us, the PRT automation is 0.10 s. Young and Stanton [57] investigated PbRT human (skilled drivers) under L1 and L2 automation by driving simulator experiments. e authors reported that the mean PbRT human of L1 and L2 to unexpected events is 3.17 s (standard deviation [sd] � 0.289 s) and 3.63 s (sd � 0.354 s), respectively. However, recent studies [58][59][60] have found that L2 automation does not worsen the brake reaction time on a statistically significant level. Naujoks et al. [61] explained that only partial automation (L2) with the instruction to keep the hands off the steering wheel leads to an increase in brake reaction times. It is important to emphasize that L2 usually allows hands-free driving only for a limited amount of time [14,62]. erefore, under the assumption of the normally distributed sample, we estimated that the 90 th PbRT human in both L1 and L2 adopts the same value of 3.50 s. us, the PbRT in both L1 and L2 is 3.60 s. e PRT ADS was found reasonably to be in the order of 0.50 s [12], larger than the PRT automation of 0.10 s due to the additional work of planning and decision-making. Din-parastDjadid [63] reviewed and investigated effects of many important factors (e.g., the time budget to collision, takeover event types) on drivers' 85 th percentile TOT value, which is an important performance metric. e authors reported that the mean and 85 th percentile TOT for the revealed stopped vehicle event is 2.34 s (sd � 1.27 s) and 3.72 s, respectively. We estimated the 90 th TOT of 3.80 s. e PbRT in L3 is the gross value of PRT ADS and TOT. Consequently, the PbRT inL3 is 4.30 s, while the PbRT in L4/5 remains the same value of 0.50 s as the PRTADS.

Deceleration Rate.
As described in Section 2.3.2, A d of all automation levels remains at its value of 3.4 m/s 2 in [25]. In addition, the L3 system usually started a preemptive braking with a conservative A d of 2.5 m/s 2 during the TOT.

Height of the Actor's Sensing Unit.
For large trucks, the recommended value of driver eye height and headlight mounting height for design are 2.33 m and 1.0 m [64], respectively, and the upward divergence angle of the light beam is 1 deg [25]. e typical placements of perception sensors on AVs are as follows: (1) forward-facing camera is located either beside the rear view mirror [49,65] or on the vehicle roof [66]; (2) Radar is installed either behind the front bumpers or on the grille [67,68]; (3) Lidar is mounted either on the roof (long range) [49] or under the grille (short or midrange) [65]. Based on the analysis in Section 2.3.3, the height of the actor's sensing unit (h s ) for L0∼L2 and L3 (during the driver fallback) remains at 2.33 m in the daytime and 1.0 m (with 1 deg upward divergence angle of the light beam) at night.
For L4/5 ADSs, the highest mounting height of sensors was selected as h s . us, without the consideration of the sensor support height, we assigned the truck height as the mounting height of Lidar, i.e., the conservative h s for L4/5 is 2.72 m [64]. Currently, Lidars integrate from 4 to 128 lasers or channels with a vertical FoV that oscillates between 20 and 45 deg [51]. Alexander et al. [69] conducted the attribute analysis of ten Lidar products of several manufacturers (e.g., Velodyne, Hesai, Ouster, and RoboSense), and presented the general specifications of each tested Lidar. Based on those studies, we selected the typical Lidar vertical FoV of 40 deg with an upward divergence of 15 deg. For L3 ADSs without the installation of Lidars, both the mounting height of camera and Radar was considered.
e generic vertical FoV of long-range Radar proposed in [70] is 5 deg. us, 2.50 m [11] (camera mounting height) and 1.0 m (radar mounting height) with 5/2 � 2.5 deg upward divergence were adopted as h s . It should be noted that human engagement in DDT fallback is still the critical scenario in L3 automation (further discussions in Section 3.1).

Height of
Object. An object height of 0.60 m for L0 − is representative of the height of passenger car headlights and taillights [25]. Due to very few findings related to h o in the context of driving automation, we assumed a minimum conservative h o of 0.6 m for all levels.

Methods
e methodology involved revising current equations of related geometric design elements by substituting matched design criteria (see Section 2.4) associated with each automation-level scenario, and then determining and comparing effects of different automation levels on those elements.
Design elements considered in this paper include the required stopping sight distance, horizontal sight line offset, and length of crest and sag vertical curve.

Evaluation Scenarios.
Based on basic driving automation levels (L0∼L5) and the analysis in Section 2, six automation levels were considered: L0 − , L0 + , L1, L2, L3, and L4/5. Additional attention was paid to the ideal scenario of L3, as L3 is the critical level where the engaged ADS initially performs the entire DDT. Some L3 features are designed to automatically perform the DDT fallback and achieve a minimal risk condition in some circumstances, such as when an obstacle-free, adjacent shoulder is presented [14]. L3 with those features was defined as ideal L3 in this paper. Because ideal L3 with Lidars is equivalent to L4/5, the scenario of ideal L3 only equipped with camera(s) and radar(s) was considered. It is easy to conclude that except for the height of the actor's sensing unit (with an upward divergence angle), values of all design criteria for ideal L3 are the same as those for L4/5. erefore, seven analysis scenarios explored in this study are as follows: L0 − (reference scenario referred to current geometric design practices), L0 + , L1, L2, L3, ideal L3, and L4/5.

Required Stopping Sight Distance.
e required stopping sight distances (SSDs) of all scenarios except L3 automation were computed as follows [25]: where V (km/h) is the design speed; i G (rise/run, m/m) the longitudinal roadway grade. Take "−" on downgrades and "+" on upgrades. Due to the preemptive braking during the TOT conducted by L3 automation functions, the required SSD in L3 can be written as follows: where A dp (m/s 2 ), 2.5 m/s 2 , is the preemptive deceleration rate activated by L3 automation. It should be noted that for all driving automation systems, the computed SSD should be within the perception range of sensors. Modern sensor systems (including camera, Radar, Lidar, etc.) allow for an environmental perception of up to 250 m.

Horizontal Sight Line Offset.
For the general safety design of the individual curve, the horizontal sight line offset (HSO) is commonly utilized to ensure that there is adequate sight distance on horizontal curves [71]. In order to study the effects of different automation levels on the sight distance across the inside of horizontal curves more conveniently, we only selected the typical horizontal sight line offset (HSO) Journal of Advanced Transportation where circular curves are longer than the sight distance for the pertinent V. erefore, the derived values of HSO proposed in [25] are: where, R (m) is the radius of the curve. It should be noted that the horizontal curve radius for the pertinent speed is determined from the superelevation rate (e) and side friction factor. e minimum radius is based on the threshold of driver/passenger comfort and that is sufficient to provide a margin of safety against skidding and vehicle rollover. Since most commercial ATs are realized by adding or refitting sensors and the onboard computer on the existing automobile model [72], there is no significant change in either vehicle size or mass. us, we assumed that R in the context of AVs maintains the same value as that in the current practice. Furthermore, given the HSO and R, the required sight distance at the upper range for a speed can be inversely computed by (3), and then that maximum allowable speed can be obtained by (1) or (2).

Crest Vertical Curves.
Except for the L3 scenario, the length of a crest vertical curve (L c ) was derived as follows [25]: where ω (%) is the algebraic difference in grades. Note that minimum lengths of vertical curves should be larger than 0.6 V (m) (also applies to lengths of sag vertical curves below) to approximate the range of current practice. In the L3 condition, the actor's sensing unit changes from the sensor (in PRT ADS ) to the driver's eye (in TOT) during the stopping sight distance (in PbRT) as the considered worst-case scenario includes the take-over process from systems to human drivers. It is obvious that the human sight line plays a salient role in the value of L c due to lower eye height and slower response time. In order not to overestimate the computed L c value in L3, it is necessary to adjust the (PbRT based) SSD in (4) to the TOT-based SSD, where (TOT based SSD) � SSD − 0.278 × V × PRT A DS . Note that although the required SSD in either L1 or L2 also includes the distance within PRT automation (0.1 s) except the human part, the overestimation in L c is not significant. Moreover, it seems to be reasonable not to separately consider SSD based on PbRT human because of the demanding immediate take-over actions by drivers who always engage in task 3 (Figure 1).

Sag Vertical Curves.
From a practical point of view, three criteria besides 0.6 V were referred above for establishing the length of a sag vertical curve (L s ): sight distance at night or undercrossings, and passenger comfort were considered.
For sag vertical curves without an overhead vertical obstruction, L s of all scenarios apart from L3 was derived as follows [25]: where β (deg) is the upward divergence angle of the sensing beam from the longitudinal axis of the vehicle. Note that passenger comfort should also be satisfied [25]: where A c (m/s 2 ), circa 0.3 m/s 2 , is the centripetal acceleration for comfortable riding.
Another practical case where L s at grade separation structures was considered. e general equation for L s at undercrossings is [26]: where h c (m) is the vertical clearance.
In the context of L3 scenario, L s computing model ( (5)or (7) maintains the same revisions as L c ( 4), i.e., SSD in both s (5) and (7) is replaced with the TOT-based SSD.

Results
Since current reliable operational design scenarios of ATs are mainly on high-type highways with high speeds and clear lane lines [73], the typical range of design speed from 60 to 100 km/h was chosen.

Required Stopping Sight Distance.
e typical i G value of 0.03 on both upgrade and downgrade was selected. e computed SSD for all scenarios on level roadways and grades were illustrated in Figure 3. As apparent from Figure 3, given the same V, the required SSDs of L1∼L3 increase and those of L0 + , ideal L3, and L4/5 decrease at any grades compared with L0 − . e required SSD in ideal L3 and L4/5 achieves the same smallest value, while that in L3 is the longest. Hence, the order of the required SSD for all scenarios is: . Moreover, the absolute value of these change rates diminishes with the higher speed or on downgrade roadways. . Figures 4(a)-4(c) show the derived values of HSO for all scenarios at the selected V of 60, 80, and 100 km/h, respectively. For the curves shown in Figures 4(a)-4(c), the end of the solid line on the curve is the minimum radius where e is equal to 12%, while the dashed portion is the values less than the standard minimum radius for a maximum e of 12%. As depicted in Figures 3(a)-3(c), given the same V, the order of HSO for all scenarios at any radii exhibited very similar patterns with that of the required SSD, i.e., L3 > L2 � L1 > L0 − > L0 + > ideal L3 � L4/5. For example, in the condition of V � 80 km/h and R � 300 m, change rates of HSO based on the reference value of 6.9 m (L0 − ) are:

Horizontal Sight Line Offset
. Similarly, the amplitudes of these change rates shrink either with higher speed or larger radius. ree cases of the given HSO and R were selected for analyzing possible impacts of driving automation levels on the maximum allowable speed (  Figure 5, the order of the allowable speed is completely opposite to that of the required HSO, i.e., L4/5 � ideal L3 > L0 + > L0 − > L1 � L2 > L3. Take Case II as an example; change rates of the maximum allowable speed are: ideal L3 and L4/5

Crest Vertical Curves.
e minimum L c for all scenarios at varied values of ω (0∼16%) to provide the required SSD for the selected V of 60, 80, and 100 km/h was displayed in Figures 6(a)-6(c). For the curves (or lines) depicted in Figures 6(a)-6(c), the short-dashed curves, dash-dotted curves, and solid curves (or lines) indicate where SSD � L c , SSD > L c and SSD < L c , respectively. As can be seen in Figures 6(a)-6(c), given the same V, the order of the minimum L c for all scenarios at any ω is basically consistent with that of the required SSD, i.e., L3 L2 � L1 ≥ L0 − ≥ L0 + ≥ ideal L3 ≥ L4/5. Note that the difference in the computed L c between L3 and L1 or L2 is so marginal that it cannot be detected in Figures 6(a)   where SSD � L s , SSD > L s and SSD < L s , respectively. In addition, values of L s on the right of the red short-dashed line can satisfy the passenger comfort criterion. Given the same V, the order of the minimum L s at any ω is in line with that of L c , i.e., L3 L2 � L1 ≥ L0 − ≥ L0 + ≥ ideal L3 ≥ L4/5. Furthermore, the order of the rounded values of K is: Take  show the minimum L s at varied ω (0∼16%) to provide the required SSD for the selected V of 60, 80, and 100 km/h. e criterion of passenger comfort was neglected because L s for all scenarios computed by (7) cannot meet that. Given the same V, the order of the minimum L s at any ω is

Discussion
e main intent of this study was to examine the impacts of varied driving automation levels on the geometric design of highways from the perspective of trucks. Particular attention was paid to explore differences in typical characteristics related to sight distances among varied levels (Section 2.3) and which design criteria of these characteristics seem reasonable and suitable for representing each level (Section 2.4). Finally, the influence of diverse automation levels with tailored features on the current models of geometric design elements (Section 3.2∼3.4) and their computed values (Section 4) were analyzed. It should be noted that the derivation of our results mainly refers to the default values in specific geometry specifications [25,26]. Results based on other guidelines where inconsistent default values may be exhibited can be easily obtained and examined by the methodology in this study. It can be anticipated that the results of this paper are sound since the differences between those values are generally small. Also, those automation features considered in this study do not include limited capacities of perception sensors in adverse weather conditions (e.g., rain or fog), i.e., the road environment for all evaluation scenarios is ideal.
ere are interesting insights, both qualitative and quantitative, to be gleaned from this study. Firstly, characteristics with respect to sight distances consisting of PbRT and h s (with β) are the main and explicit differences among manifold automation levels. Previous studies [10,12] have investigated those characteristics in the context of highly automated driving. We extended previous findings to all driving automation scopes (L0∼L5) supported by a broad body of literature. Notably, decomposed components of PbRT for each level were established. e PRT automation and the PRT ADS followed by TOT were introduced into the PbRT structure in L1/L2 and L3, respectively, where the control right of ATs is shared between the human driver and systems to some extent. Furthermore, sensing actors for all levels were reconsidered and adjusted in the worst-case scenario. In short, features of perception sensors and system functions will replace those of humans gradually in defining characteristics of sight distances.
Another important finding of this study is that with the advancement of automation levels, requirements of geometric design elements including the required SSD, HSO, and the minimum L c and L s moderate initially (L0 + ), then develop toward the strict direction (L1∼L3), and finally diminish significantly (L4/5) compared with those in L0 − , for e.g., longer required SSD is necessitated in L1∼L3 and smaller L c in L0 + or L4/5. In other words, both L0 + and L4/5 (especially the latter) can fully adapt to the geometry condition of highways in service, while L1∼L3 (especially L3) need higher-type geometric specifications to ensure safety [73]. e explanation for this finding could be that quick PbRT or/and high h s (with large β) are achieved in L0 − and L4/5, while the slower PbRT human or TOT and the continued request of drivers in the control loop deteriorate the anticipated performance of automation systems in L1∼L3.
is study also found that larger maximum allowable speeds can be obtained in both L0 − and L4/5 (especially the latter) but smaller in L1∼L3 (especially L3) given the HSO and R. e latter (smaller allowable speeds in L1∼L3) can be verified by Garcia et al. [74] through a real semiautonomous vehicle test along highways. A more interesting finding is that the requirements of all considered design elements in  Some limitations of the present study must be considered. Obviously, a single value selected for each characteristic such as PbRT or h s cannot provide a complete description of it. More validations and research have to be conducted before definitive conclusions of design criteria for all automation levels can be drawn. However, even at this stage, the design criteria estimated in this article seem to be in a reasonable agreement with the data exhibited in published studies and released products. It seems reasonable to expect that this approach will help to solve practical problems such as the geometric design of highways more economically and conveniently. Also, this study was conducted from the perspective of trucks, which covers only one class of design vehicles. e passenger car is definitely one of the most representative vehicle types for on-road motor vehicles equipped with driving automation systems and for purposes of geometric highway design. Results in the context of passenger cars can be easily obtained and examined by the methodology in this study. Another limitation of this study is that the typical operation form of ATs, platoon, or roadtrains is not considered. Although one single truck in this study can be regarded as the leader in the platoon, further research is need on actual longitudinal and lateral driving behaviors of followers under a real road environment. Finally, this study only estimated the sight distance on the horizontal or vertical plane and overlooked the threedimensional (3D) nature of the highway infrastructure. e authors have proposed an automatic framework for 3D highway sight distance analysis in [71]. Detailed impact analysis of automation systems on sight distance along the 3D highway alignment is expected in the following work.

Conclusions
e results presented here offer new insights into differences in typical characteristics related to sight distances among varied driving automation levels and possible positive or negative effects of those characteristics on geometric highway design. Results show that high or full driving automation could substantially lower the requirements of geometric design, which is consistent with previous findings. Active safety systems have a similar role but with moderate effects. Differently, the driver assistance and partial or conditional automation systems put a higher demand on the road geometric design from the safety point of view.
Primary implications of this study are expected to have design consequences for the real-world geometry of dedicated lanes. Also, our findings can be used to facilitate the identification of potential risk sections on highways in service for the operation of L1∼L3 and otherwise provide a methodological basis for the formulation of speed limit policy. Furthermore, this paper may be useful in urging researchers or technicians of automated driving systems to take into account driver states and performances in the takeover process.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.