Framework for Assessing Potential Safety Impacts of Automated Driving Systems (2022)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

B-1   This appendix provides analytical examples for applying the framework to three prominent ADS features that are envisioned, namely, conditional traffic jam assist, highway platooning, and fleet-operated automated driving system–dedicated vehicle (ADS-DV). The following sub- sections apply the framework steps described in Chapter 3 to assess the safety impacts of the ADS features. ADS Feature: Conditional Traffic Jam Assist Understanding the ADS Feature Feature Description Level 3 conditional traffic jam assist (TJA) is a subsystem of adaptive cruise control (ACC) with extended lane-keeping assist functionality. The types of sensors needed by this feature and the feature’s operating settings vary significantly depending on the offering manufacturer. However, the core of the different TJA systems operates on the same concept. Typically, the automated vehicle (AV) uses onboard sensors such as cameras, radars, and lidars to identify slow- moving traffic; once slow traffic is identified (less than 35 mph), the TJA engages. The AV then locks on to the vehicles in front within the same lane and handles the driving task such that the TJA-equipped AV stays about 3 seconds behind the vehicle ahead of it. AVs equipped with TJA can assume certain steering tasks over a speed range up to 35 mph on roads that are in good condition, as long as the traffic is moving slowly. The system uses the AV sensors to guide the vehicle by making gentle steering movements within system limits and orients itself to lane markings, roadside structures, and other vehicles on the road. For this fea- ture, a human operator is the fallback for the dynamic driving task. When TJA operates beyond its operational design domain (ODD) or system limits (see “Operational Design Domains” in Chapter 3 for details), such as when the traffic speed is above 35 mph; lane markings are not clear; or there is a sharp curve ahead, inclement weather, or poor lighting, the driver must assume driving tasks again. If the driver does not, the system warns the driver in several stages, and as a final measure, it would engage an emergency assist, bringing the car to a safe stop. TJA perfor- mance can be affected by inclement weather, poor lighting, and poor infrastructure conditions. Expected Market The Level 3 TJA was supposed to launch in the United States starting in 2019. A few years earlier, Audi announced the inclusion of a Traffic Jam Pilot driver-assist technical feature in the 2019 Audi A8 luxury sedan. Although Audi might be able to make this commitment to the European automotive market, Audi could not offer the feature in any of the 2019 U.S. models, primarily due to legislative barriers and some infrastructure challenges. The main infrastructure A P P E N D I X B Additional Example Scenarios

B-2 Framework for Assessing Potential Safety Impacts of Automated Driving Systems obstacle is the discrepancy among U.S. states in lane markings, signage, traffic signals, and road configuration. For instance, pavement markings, particularly skip lines, vary in width and length among states, making it difficult to develop a reliable system that is able to function accurately across the entire United States. As this feature will be offered at the Level 3 automation level, drivers engaged in road jams are no longer required to keep their hands on the steering wheel or continuously monitor the vehicle and the road. They must only be alert and prepared to take over the task of driving when the system prompts them to do so. Unlike previous AV features offered (Tesla Autopilot, Cadillac Super Cruise, or Nissan Pro Pilot) where drivers are instructed to watch the road, this one does not require this. This raises concerns about whether the driver would be able to react promptly to the situation and take full control of the vehicle. Expectations of full and complete attentiveness of the driver in this model are unlikely. Hence, insurance companies might offer higher-than-average premiums for these vehicles. Audi mentioned concerns with the lack of clear federal regulations for autonomous driving technology and the absence of comprehen- sive federal standards. Audi believes state regulations make it impossible to sell these products nationwide. Several car manufacturers (Volvo, Ford, and Google, among others) expected the safety and policy concerns associated with Level 3 automation and decided to skip it and jump directly to Level 4 and Level 5 automation. In fact, Ford announced in 2018 that the company is finalizing its own TJA; however, Ford has offered no timeline for the debut, which is expected to be on a Level 4 vehicle. Recently, Ford began investigating stepping back to the Level 3 automation, with the aid of cameras and other systems that can ensure drivers are paying attention at the wheel (Martinez, 2019). Researchers and drivers may argue that coupling the Level 3 features operation with the driver’s continuous attention is contradicting the SAE definition (SAE, 2018) for Level 3 automa- tion, where the driver is not required to be attentive to the driving task once the ADS is in control. In this regard, the market of the Level 3 does not seem to be promising for expanding the deployment of ADS features. Specifically, all ADS features expected to launch on U.S. roads at Level 3 automation are susceptible to safety, policy, and regulations barriers. Although the policy and regulations barriers hold in the subsequent automation levels (4 and 5), it is anticipated that resolving these issues can be easier for highly automated vehicles (HAVs) (Levels 4 and 5). Further, the global TJA market (all automation levels) was estimated at $1.08 billion in 2016 and is expected to reach over $48 billion by 2026 (BIS Research, 2017). The baseline status of the automation market implies that the growth of the TJA market is likely to occur mainly at automation Level 2 until HAVs (Levels 4 and 5) are offered for commercial deployment. On the basis of these facts, and extrapolating from the forecasted market size for the Level 3, 4, and 5 vehicles until 2030 (McKinsey & Company, 2016; Frost and Sullivan, 2018), the project team projected the deployment scenarios for this feature to follow a low disruption pattern. Table B-1 lists a few examples of the TJA features that are commercially deployed. According to currently available information, none of the car manufacturers offered their TJA feature as pure Level 3 in the feature’s description in the user manual. In 2019, Audi reverted to promoting its TJA as a subsystem of ACC. For all of these features, car manufacturers are obligating drivers to stay 100% attentive to the roadway and to ensure that these features are operating within their specified ODD. Anticipated Deployment Scenario Operational Design Domain Table B-2 summarizes the anticipated ODD elements of the conditional TJA. As discussed earlier, the ODD elements are predicted for two timelines, the short term (next 5 years) and

Additional Example Scenarios B-3   the medium term (next 5 to 10 years). In addition, the table outlines the major deployment specifications envisioned for the feature that are expected to affect the safety assessment of the feature. The deployment elements are also provided for the two timelines. Table B-3 provides a summary of the technology specifications and key infrastructure needs pertinent to the envisioned TJA scenarios. As shown in Table B-3, the vehicle is expected to rely on different generations of sensor packages during the different planning horizons. Typically, Gen I is envisioned to persist during the short-term planning horizon while Gen II of the sensor package is expected to take over during the medium-term planning horizon. A detailed description of each of the anticipated scenarios, in terms of technology and infrastructure needs, follows. Stage of Technology Development • First Generation (Gen I): This is the first version of TJA technology and sensor package and their underlying computational algorithms for processing the data. Typically, this pack- age embraces the needed combination of sensors such as forward-facing cameras, radar, ultrasonic sensors, laser scanners, and inertial measurement unit (gyroscopes and acceler- ometers) with a priori digital maps (lane-level detail). This package will not have a good object detection capability in low-visibility conditions, limiting the ODD to certain conditions (e.g., light rain, no snow, good lane markings). • Similarly, the underlying perception algorithms for processing the data are primitive and are capable of handling the basic computations needed for the proper functionality of the TJA feature within the ODD and specified deployment context. Car Manufacturer Traffic Jam Assist Feature Audi The closest production car to that reality was to be the new Audi A8, equipped with the new Traffic Jam Pilot feature capable of navigating highway gridlock while the driver performs secondary tasks (not relevant to driving), but regulatory hurdles forced Audi to sideline the system for US shoppers, at least for now. Currently, Audi is offering its TJA as a subsystem of the ACC. TJA can assume certain steering tasks over a speed range up to 35 mph on roads that are in good condition, as long as the traffic is moving slowly. The base ACC system accelerates and brakes to keep the vehicles at the desired distance from the vehicle ahead. BMW BMW’s ACC that operates all the way to a stop goes by Active Cruise Control with Stop & Go™. One step beyond that, Active Lane Keeping Assist with Side Collision Avoidance constitutes hands-on, lane-centering steering that can work down to a stop in certain traffic conditions. Finally, Extended Traffic Jam Assistant (ETJA) is a new feature available on the redesigned 3 Series and X5 plus the new 8 Series and X7. ETJA enables hands-free driving at low speeds on divided highways as long as the driver is paying attention, something the car intuits with a driver-facing camera. Mazda Mazda Radar Cruise Control with Stop and Go. The redesigned Mazda 3 has a TJA system with low-speed lane-centering steering, but Mazda did not offer the feature in the US market for 2019. Acura Depending on the car, AcuraWatch includes a Lane-Keeping Assist System, ACC with Low-Speed Follow, or both. Meanwhile, the RLX and RLX Sport Hybrid add TJA that incorporates lane-centering steering down to a stop. Table B-1. Examples of commercially deployed TJA features (Source: Mays, 2019).

B-4 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Timeline Operational Design Domain Level Additional Deployment Context Short term (low disruption) • Market share = 2% • Fleet share = 1% • VHT share = 1% • Urban and rural highways (divided). • Intrastate and freeways (between urban or rural areas). • Operating only in traffic conditions where traffic speed is below 35 mph. • Operating only in clear and good weather condition (e.g., no rain, snow, etc.). • This is a conditional automation Level 3 feature where the driver is the fallback for the dynamic driving task (DDT). • Operating with mixed traffic. • Operating in stop-and-go traffic at headway of about 3 seconds (2 to 4). • No lane changes. • Not operating on off- or on-ramps. • Not operating on horizontal curves. • Not operating near intersections (signalized or stop controlled). Medium term (low disruption) • Market share = 6% • Fleet share = 3% • VHT share = 4% • In addition to the short-term ODD elements: Urban one-way arterial streets. Urban and rural multilane arterial streets (divided). Operating in traffic conditions where traffic speed is below 50 mph. Could operate in light and moderate rain conditions. • This is a conditional automation Level 3 feature where the driver is the fallback for the DDT. • Operating with mixed traffic. • Operating in stop-and-go traffic at headway of about 3 seconds (2 to 4). • Can perform simple lane change maneuvers. • Could navigate through work zones. • Could operate on off- or on-ramps. • Could operate on horizontal curves. • Could navigate through intersections. Table B-2. TJA L3 deployment scenarios of interest. Planning Horizon Vehicle Type Sensor Package Key Infrastructure Requirements Digital Physical Short term (low disruption) Light duty. Gen I. Global Positioning System (GPS)-relevant infrastructure. • Pavement markings in excellent condition. • Signage in excellent or good condition. • Highway median barriers. Medium term (low disruption) Light duty. Gen II. • GPS-relevant infrastructure. • Vehicle-to-vehicle (V2V)- and vehicle-to- everything (V2X)- relevant infrastructure. • Weather data. • Lane and pavement markings in excellent condition. • Signage in excellent or good condition. • Work zone data. • Dynamic traffic data. Table B-3. Sensor specifications and infrastructure requirements of TJA scenarios.

Additional Example Scenarios B-5   • These algorithms are in early development and may have more errors than later, more mature technology, leading to lower safety performance and/or lower percentage of time operating in autonomous mode (high disengagement rate). Sensors and computation algorithms used at this stage are commercially available and currently operate in certain vehicles. • Second Generation (Gen II): Gen II is of a more advanced stage of technology than first- generation models or systems. In addition to the sensor types included in Gen I, this gen- eration embraces high-fidelity lidar sensors and a vehicle onboard unit (OBU). All Gen II sensors are newer, more advanced, accurate, and have longer perception range than the Gen I sensors. A key feature of the Gen II sensor package is providing layers of redundancy to one another. The perception algorithms frequently cross-check the data from different sensors to improve the chance that no object is left undetected and to eliminate false positives. • This sensor package manifests as a mature sensor fusion technology that is able to combine the sensing capabilities of multiple sensors, resulting in more reliable and robust perception with a broad sensing scope. To this end, the underlying perception algorithms for processing the data are more advanced and are capable of performing complex sensor fusion calculations, enabling the operation in expanded ODD and deployment context (check). • Another envisioned key feature of this generation of sensors is integrating V2V and vehicle- to-infrastructure (V2I) communication within the vehicles through the OBU. This would provide opportunities for the vehicle to receive real-time dynamic data for weather, work zones, and traffic. Table B-4 provides a qualitative assessment of the technology state of the different sensor package generations. The table highlights the key functional and technical differences between the two generations as well as the operational atmospheric conditions. Infrastructure Needs and Impacts • Short-Term Planning Horizon: From the infrastructure perspective, the short-term planning scenario for this feature has some key limitations, such as depending on surrounding vehicles and having clear roadway edges. There are also potential challenges around changes to the roadway and being unable to navigate work zones, horizontal curves, broken-down Sensor Package Scenario Technology State Comparison Qualitative Assessment of Technology State Operational Atmospheric Conditions Key Functional and Technical Differences Gen I • Lower-priced vehicles. • Less sophisticated algorithms, making driverless mode active less often. Weather: clear, wind. All Gen I features are offered in the Gen II package with enhanced capabilities. Gen II • Higher-priced vehicles. • More sophisticated algorithms, making driverless mode active more often. • Communication with vehicles and roadway infrastructure. Weather: clear, wind, light, and moderate rain conditions. • Lidar. • V2X communications. • V2V communications. Table B-4. Scenario technological specifications of traffic jam assist.

B-6 Framework for Assessing Potential Safety Impacts of Automated Driving Systems vehicles, or poor weather conditions. The dependence on surrounding cars to help navigate (i.e., the requirement to have a car in front and cars to the side) decreases the dependence on sensors for lane detection. Still, the Level 3 feature must detect the lanes and roadway edge based on lines, barriers, guard rails, or other markings. To improve functionality of these sys- tems, infrastructure owners and operators (IOOs) may wish to inventory and maintain clear roadway edges and lane assignments. • Medium-Term Planning Horizon: Digital infrastructure pertaining to connectivity plays an integral role in expanding the capabilities of this feature. The feature may benefit from real-time message updates for weather, traffic, and other changing conditions. In heavily congested areas, IOOs may consider investments for connectivity. Additionally, connectivity such as V2I communications could provide vital information to the vehicle, such as work zones and stopped vehicles. For example, roadside equipment in Virginia’s connected corridors along portions of I-66, I-495, U.S. 7, U.S. 29, and U.S. 50 provides data flows for live transmission of incidents, weather, work zone, and variable and dynamic message signs that may benefit Level 3 features. Risk Assessment Risks • The major risk for this feature resides in the known challenges for ensuring a safe transi- tion of the DDT task from machine to human. One such concern is the poor understanding of how much time and context a human driver needs to regain control and proper situational awareness. Another challenge is monitoring driver state. These challenges impose extra safety and liability risks at the Level 3 automation—compared to all other levels of automation— due to human factors. The conservative approach implemented by OEMs so far for handling these challenges is promoting the feature as a hybrid automation of Level 2 and Level 3. In these hybrid automation levels, the human driver is liable at all times, including when the ADS is handling the DDT. • Another concern with the implementation of ADS features (Level 3 and Level 4) is the expected degradation of some driving skills—depending on the ADS feature—due to reliance on automation and lack of exposure to DDT. Skill degradation is not limited to psychomotor skills (i.e., those pertaining to performing maneuvers), but also include deci- sion-making skills. • A potential risk envisioned with the deployment of TJA is that it is not expected to have one uniform TJA system in place across all OEMs and car models. Therefore, drivers may need to take caution when using a new ADS-equipped vehicle and becoming familiar with the machine-user interface and operation specifications of the new TJA system. For instance, not all TJA systems would use the same headway when following a leading vehicle. A driver familiar with one system could fail to react safely for a request-for-control takeover by another system. This failure may not only be due to the difference in default operation specifications but could be attributed to human factors related to different machine-user interfaces for control takeover requests. States could consider updating their theoreti- cal content for state driving handbooks to educate new applicants and ADS users about these issues. Similarly, TJA systems may also not work properly in certain weather condi- tions like heavy rain or fog, if there is mud or snow on the sensors, or if roads are slippery. Drivers should be fully aware of the limitations of the system and capable to interact with the machine interface efficiently. • At low market penetration rates, TJA systems could contribute to new crash types since they may have negative safety impacts on the nearby traditional, nonautomated vehicles. For example: – Traditional nonautomated vehicles following TJA-equipped vehicles could experience increased risk for rear-end crashes at low penetration rates.

Additional Example Scenarios B-7   – At low penetration rates, TJA-equipped vehicles could result in more aggressive and fre- quent lane-change maneuvers by following, nonautomated vehicles. This could increase the crash risk for sideswipe crashes and other crashes in the aggregate traffic stream due to disruptions in traffic flow. Opportunities • TJA has the potential to reduce the number of crashes, such as rear-end collisions, occurring on congested roads due to human driver fatigue. TJA is a subsystem of ACC with extended lane- keeping assist functionality. Therefore, these technologies also have the potential to decrease crashes occurring on congested roads pertaining to lane-change maneuvers. Overall, these technologies are envisioned to improve safety on roads during congested traffic conditions. • TJA is among the ADS features that provide massive opportunities for better fuel savings and longer vehicle life. Automating the vehicle in congested conditions improves the vehicle’s fuel economy and motor lifetime. Recipients of funding from the U.S. Department of Energy’s NEXTCAR Program are commercializing these types of technologies (ARPA-E, 2020). • With increased market share, TJA-equipped vehicles provide the potential for implementing innovative eco-driving and speed harmonization techniques when such vehicles are con- nected (supporting V2V or V2I communication systems). More sophisticated eco-driving and speed harmonization techniques could be considered, such as geofencing of a cordon that may be scalable and movable. Goals and Hypothesis For the TJA deployment scenario, the safety objectives and hypothesis for how procedures and processes can influence these outcomes in future deployments are formulated in terms of a goal and hypotheses. The goal of TJA is to reduce the frequency and severity of crashes across a transportation agency’s system (e.g., state of California or city of Charlotte) through the adoption of TJA features (SAE Level 3) and supporting infrastructure. The overall hypothesis is that TJA will improve safety on urban divided streets across a state by reducing rear-end crashes within 3 to 5 years. The expected change in the number and percentage of rear-end crashes will depend on market penetration, which is explored in the analysis. The questions to evaluate this hypothesis are listed below and relate to crash types, crash severity levels, infrastructure, and data. 1. How will the frequency of certain types of crashes (e.g., rear-end crashes due to human error, including driver fatigue, inattention, and inebriation) change in relation to safety? 2. How will the severity of certain crash types change? While it is anticipated most crashes affected would be property damage only (PDO), the feature could reduce the severity of crashes that do occur (e.g., reducing possible injury crashes to PDO crashes). 3. Will safety change if the ODD is extended in which TJA features can operate? For example, it is hypothesized that TJA features do not perform well on facilities with poor pavement markings, and there is an opportunity to further improve safety through infrastructure invest- ments in lane markings (application of new markings and maintenance of existing markings). 4. Will safety and deployment of TJA features change if work zone data are shared that improve the safety performance of the ADS and make it easier for manufacturers to deploy the tech- nology ubiquitously? In summary, through the deployment of TJA, it is hypothesized that certain crash types will be reduced, crash severity will lessen, infrastructure improvements will help advance and extend the deployment and use of TJA, and data sharing can help improve safety performance. However, certain crash types might increase (e.g., sideswipe crashes).

B-8 Framework for Assessing Potential Safety Impacts of Automated Driving Systems While this provides an overview of potential safety impacts, it is important to perform a crash sequencing exercise to think through the contributing factors and precipitating events that lead to a crash. The following are a few examples related to TJA: 1. Rear-end: a. Congested conditions cause vehicle 1 to stop. Driver of vehicle 2 is distracted, falls asleep, or is otherwise inattentive. Vehicle 2 rear-ends vehicle 1. b. Congested conditions cause vehicle 1 to stop. Driver of vehicle 2 is fully attentive but following too closely whereby vehicle 2 is not able to stop in time. Vehicle 2 rear-ends vehicle 1. c. Congested conditions cause vehicle 1 to stop. Driver of vehicle 2 is fully attentive and adverse weather contributes to limited stopping distance whereby vehicle 2 is not able to stop or slow. Vehicle 2 rear-ends vehicle 1. 2. Sideswipe: a. Driver of vehicle 1 is distracted, falls asleep, or is otherwise inattentive. Vehicle 2 is traveling in the same direction in the adjacent lane to vehicle 1. Vehicle 1 drifts into the adjacent lane and sideswipes vehicle 2. b. Vehicle 1 is equipped with TJA. Vehicle 2 is not equipped with TJA and is follow- ing vehicle 1. Vehicle 3 is traveling in the same direction in the adjacent lane to vehicle 1 and vehicle 2. Driver of vehicle 2 becomes inpatient with steady speed of vehicle 1. Driver of vehicle 2 abruptly changes lanes to pass vehicle 1 and does not see vehicle 3 in the blind spot. Vehicle 2 sideswipes vehicle 3. From the anticipated capabilities of vehicles equipped with TJA and the above crash sequencing, the project team identified specific opportunities for TJA to mitigate crashes. For example, TJA is not expected to operate in adverse conditions, so there is limited potential to mitigate crashes related to sequence 1c; however, TJA is expected to provide opportunities to miti- gate crashes related to sequence 1a and 1b. TJA could also contribute to (or be associated with) crashes related to sequence 2b. Method of Analysis Components of the Analysis To estimate the safety effects of TJA, an analysis to quantify its impacts needs to be performed using crash data and the types of crashes TJA is expected to influence. As previously described, the ODD includes divided roadways in urban environments. TJA can only operate in clear weather conditions (clear or windy) for the short-term timeline and in clear (clear or windy) with moderate rain weather conditions on roads with well-maintained lane markings. Also, TJA can operate in conditions where there is a lead vehicle and operating speeds are a maximum of 35 mph. Therefore, the focus facility type is defined as all urban divided roads with any speed limit, with the assumption that the feature and resulting safety impact are relevant only during congested conditions (e.g., peak time periods). Congested conditions are assumed to occur from 6:00 to 9:00 a.m. and 4:00 to 7:00 p.m. TJA will potentially affect only certain crash types. The main crash type expected to decrease is rear-end crashes. This is due to the ability of TJA to keep a safe distance between it and the vehicle in front and brake when needed. Differently, sideswipe crashes could potentially increase. The rationale is that if a vehicle using TJA is maintaining a safe following distance or traveling at a slower speed, a trailing vehicle might see it as an opportunity to switch lanes and increase the potential for a sideswipe crash. While only certain crash types are expected to change, all crash severity levels are expected to be affected by TJA. Related PDO crashes could be eliminated, and crashes of higher severity levels could decrease in severity. A summary of the components of the analysis is shown in Table B-5.

Additional Example Scenarios B-9   To estimate the safety benefits or disbenefits of TJA, technical assumptions about the scenario were made in analyzing crash data. An “optimistic scenario” is assumed where there is a maxi- mum crash reduction. The scenario assumes a low disruption of vehicles with TJA, for example, a certain percentage of vehicles with the feature within 5 years. Other technical assumptions relate to the crashes that are expected to be affected by TJA. It is assumed that TJA will not eliminate all rear-end crashes in the ODD described above where the trailing vehicle is ADS equipped and the feature is activated. The assumption is that TJA will eliminate only a certain percentage of crashes where the trailing vehicle is ADS equipped and the feature is activated. Varying penetration and activation rates are assumed to account for different scenarios and number of vehicles equipped with TJA. ADS-equipped vehicles may also be involved in other crash types when the driver takes control (e.g., sideswipe crashes when a driver changes lanes). The expected timeframe for achieving success depends on the expected timeline for deploy- ment and penetration rates of ADS-equipped vehicles. In this scenario, the expected timeline is assumed to pertain to short-term planning, which is a fast uptake (3 to 5 years). Analysis Methods Data Sources. The main source of crash and roadway data used to estimate the potential effect of TJA on safety was the Highway Safety Information System (HSIS). The HSIS dataset consists of crash and roadway data from seven states and one city (California, Washington, Minnesota, Illinois, Ohio, North Carolina, Maine, and Charlotte). The data can be requested online and can be requested for specific states and variables needed. For this analysis, data for Charlotte were obtained because of its urbanization, availability of variables, and availability of supplemental datasets that can be merged with HSIS data. The raw HSIS data consist of roadway, crash, vehicle, and occupant files that can be merged. Data include information about the roadway (number of lanes, divided or undivided, speed limit, etc.), crash (day, light condition, weather, contributory factor, etc.), vehicle (vehicle type, etc.), and occupant (age, condition, etc.). Component Description Functional classification. Freeways and expressways. Arterials. Major and minor collectors. Area type. Urban. Speed conditions. Any speed limit, congested conditions (weekdays 6:00 to 9:00 a.m. and 4:00 to 7:00 p.m.). Number of lanes and road configuration. Multilane, divided. One or more lanes, one-way. Weather conditions. Clear, wind, moderate rain. Crash types. Rear-end, sideswipe. Crash severity level. All severity levels. Table B-5. Components of the analysis.

B-10 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Roadway data from Connect NCDOT (Road Characteristics Arcs; https://connect.ncdot.gov/ resources/gis/Pages/GIS-Data-Layers.aspx) were used to supplement the HSIS roadway data and provide additional variables for analysis, including road functional class. Other supplemental data included census information from Charlotte’s Open Data Portal (https://data.charlottenc. gov/). These data include household income information. The census data can be used to estimate the prevalence of ADS features based on median household income in an area. However, to do so, assumptions need to be made about income and the associated possibility of owning an ADS-equipped vehicle. Both supplemental databases were merged with the HSIS roadway and crash data. Define Metrics. The primary metric for assessing the safety impact of TJA is the change in crashes by frequency and severity. In particular, the change in crashes is based on crashes of interest for the deployment scenario and crashes that TJA has the potential to impact, described in the “Components of the Analysis” section. The crashes of interest are identified by categorizing and filtering the crashes based on variables selected to identify crashes with the potential for TJA to reduce. Most were identi- fied in the “Components of the Analysis” section, which identified several variables that can be used to filter based on the ODD and deployment context. Crash types include rear-end and sideswipe. The weather conditions need to be clear or windy for the short-term timeline and clear, windy, or moderate rain for the medium-term timeline. Other conditions that need to be considered if information is available include visible lane markings (maintained and clear of snow or debris) and visible signage (high retroreflectivity). If this information is not readily available, then assumptions need to be made based on local knowledge, judgment, or a sample of data. In this case, lane marking and signage condition information is not available in the Charlotte database, and assumptions are made with respect to those factors. Evaluation Method. It is important to select an appropriate method to test the hypoth- esis and related questions based on the available data. This evaluation uses crash history and identifies the crashes that could potentially be affected by TJA. The evaluation estimates crash reduction by frequency and severity. Table B-6 lists the categories of crashes used to estimate the impact of TJA. Category Crash Type Crash Severity 1 Rear-end PDO. 2 Rear-end Possible injury. 3 Rear-end Suspected minor injury. 4 Rear-end Fatal and suspected serious injury. 5 Sideswipe PDO. 6 Sideswipe Possible injury. 7 Sideswipe Suspected minor injury. 8 Sideswipe Fatal and suspected serious injury. Table B-6. Crash types and severity levels to estimate TJA impact.

Additional Example Scenarios B-11   A sensitivity analysis is used to explore various assumptions related to penetration rates and probabilities that the trailing vehicle has TJA, has it activated, and it functions properly. The change in crashes is calculated by subtracting the crashes TJA can reduce or increase from the total number of crashes for the given years. This is repeated for each crash type and severity level. The following equation shows the change in crashes as a percent change, where a positive percent change indicates a safety benefit, and a negative change indicates an increase in crashes. ( )( ) ( ) = − −        ×Percent change in crashes 1 100 Total crashes Crashes impacted by ADS feature Total crashes Eq. 7 Results HSIS crash and roadway data were merged with road data from Connect NCDOT and census data from Charlotte’s Open Data Portal to identify crashes that could be impacted by TJA for 6 years (2012–2017). First, total crashes on the road types of interest were identified. Next, crash types that could specifically be impacted by use of TJA were identified. The total crashes were categorized based on the variables of interest and filters listed in Table B-6. Three different scenarios for categorizing the Charlotte crash data were used for the analysis to identify the crashes that could potentially be affected by the use of TJA. The scenarios include: 1. Urban, multilane, divided roadways (non-freeways and expressways) during peak periods. 2. Urban, one-way roadways (non-freeways and expressways) during peak periods. 3. Divided, multilane freeways and expressways during peak periods. Total Crashes on All Road Types Table C-1 through Table C-5 in Appendix C display crash statistics for all crashes in Charlotte (on all road types). The crashes are separated by severity level (Table C-1), crash type (Table C-2), weather condition (Table C-3), speed limit (Table C-4), and speed limit and weather condition combined (Table C-5). Urban, Multilane, Divided Roadways (Non-freeways and Expressways) During Peak Periods The total crashes were categorized to identify those that TJA has the potential to impact. The first categorization included separating out urban, multilane, divided roads (non-freeways and expressways) during congested conditions (weekday peak periods). It also included crashes occurring during clear or windy weather conditions, all non-clear and non-windy weather con- ditions, and rain weather conditions. Table B-7 displays total crashes by severity level on these roadways. Table B-8 includes only rear-end crashes on these roads of interest, and Table B-9 includes only sideswipe crashes on these roads of interest. Urban, One-Way Roadways (Non-freeways and Expressways) During Peak Periods The second categorization of the total crashes separated out urban, one-way roads (non-freeways and expressways) during congested conditions (weekday peak periods). It also included crashes occurring during clear or windy weather conditions, all non-clear and non-windy weather conditions, and rain weather conditions. Table B-10 displays total crashes by severity level on these roadways. Table B-11 includes only rear-end crashes on these roads of interest. Table B-12 includes only sideswipe crashes on these roads of interest.

B-12 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Fatal and suspected serious injury 16 5 4 Total 4,256 582 541 Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 2 1 1 PDO 2,890 404 371 Possible injury 1,172 152 146 Suspected minor injury 176 20 19 Table B-7. Total crashes on urban, multilane, divided roadways during peak periods (2012–2017). Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 2 1 1 PDO 1,537 250 235 Possible injury 783 111 108 Suspected minor injury 84 9 9 Fatal and suspected serious injury 6 1 1 Total 2,412 372 354 Table B-8. Rear-end crashes on urban, multilane, divided roadways during peak periods (2012–2017). Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 740 69 67 Possible injury 128 13 12 Suspected minor injury 8 0 0 Fatal and suspected serious injury 2 0 0 Total 878 82 79 Table B-9. Sideswipe crashes on urban, multilane, divided roadways during peak periods (2012–2017).

Additional Example Scenarios B-13   Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 13 1 1 Possible injury 2 0 0 Suspected minor injury 0 0 0 Fatal and suspected serious injury 0 0 0 Total 15 1 1 Table B-10. Total crashes on urban, one-way roadways during peak periods (2012–2017). Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 2 0 0 Possible injury 2 0 0 Suspected minor injury 0 0 0 Fatal and suspected serious injury 0 0 0 Total 4 0 0 Table B-11. Rear-end crashes on urban, one-way roadways during peak periods (2012–2017). Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 7 1 1 Possible injury 0 0 0 Suspected minor injury 0 0 0 Fatal and suspected serious injury 0 0 0 Total 7 1 1 Table B-12. Sideswipe crashes on urban, one-way roadways during peak periods (2012–2017).

B-14 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Divided, Multilane Freeways and Expressways During Peak Periods The third categorization of the total crashes included separating out urban, divided, multilane freeways and expressways during congested conditions (weekday peak periods) (Tables B-13, B-14, and B-15). It also included crashes occurring during clear or windy weather conditions, all non-clear and non-windy weather conditions, and rain weather conditions. Impact of TJA The crashes with potential for impact from TJA were further reduced to estimate a quantifi- able number of crashes impacted. The potential reduction in crashes is based on the vehicle miles traveled (VMT) share described in the deployment scenarios. The VMT shares for different timelines of TJA are: Short-term (low disruption) VMT share = 1% Medium-term (low disruption) VMT share = 4% Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 286 27 29 Possible injury 167 19 19 Suspected minor injury 10 0 0 Fatal and suspected serious injury 0 0 0 Total 463 46 48 Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 416 48 44 Possible injury 199 29 29 Suspected minor injury 14 1 1 Fatal and suspected serious injury 0 0 0 Total 629 78 74 Table B-13. Total crashes on urban, divided, multilane freeways and expressways during peak periods (2012–2017). Table B-14. Rear-end crashes on urban, divided, multilane freeways and expressways during peak periods (2012–2017).

Additional Example Scenarios B-15   In addition to the percentage of VMT share shown above, other percentages of VMT share were used as a sensitivity analysis to account for two variants: (1) vehicles that have the ADS but not engaged, and (2) vehicles that have the ADS available and engaged but it does not operate as intended. Given the wide range of projections for market penetration, 50% of the VMT shares (listed above) were assumed as the values for the sensitivity analysis, listed below. This helps to show the relative difference in results if only half of the expected market penetration is achieved. Short-term VMT share for sensitivity analysis = 0.5% Medium-term VMT share for sensitivity analysis = 2% The following sections display the short- and medium-term impact of TJA. Rear-End Crashes Short-Term Timeline. The ODD for the short-term timeline includes urban, divided highways and freeways, where traffic speed is below 35 mph and weather conditions are clear or windy. The potential crashes impacted per year are then reduced using the VMT percentages for different sensitivities. Table B-16 and Table B-17 show the expected reduction in crashes per year by severity for the short-term timeline. The percent crash reduction with respect to average total crashes per year is calculated below using a VMT share of 0.5%. = − −       × =Percent crash reduction per year 1 23,975 2.396 23,975 100 0.01% The percent crash reduction with respect to average total crashes per year is calculated below using a VMT share of 1%. = − −       × =Percent crash reduction per year 1 23,975 4.792 23,975 100 0.02% Rear-End Crashes Medium-Term Timeline. The ODD for the medium-term timeline includes all elements from the short-term timeline with the addition of one-way roadways and including crashes during rain (Tables B-18 and B-19). Severity Level Number of Crashes in Clear Weather Conditions Number of Crashes in Other Weather Conditions Number of Crashes in Rain Weather Conditions Blank 0 0 0 PDO 95 7 6 Possible injury 16 2 2 Suspected minor injury 1 0 0 Fatal and suspected serious injury 0 0 0 Total 112 9 8 Table B-15. Sideswipe crashes on urban, divided, multilane freeways and expressways during peak periods (2012–2017).

B-16 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Severity Total Crashes on All Roads and in All Weather Conditions (2012–2017) Rear-End Crashes on Urban, Multilane, Divided Roadways during Peak Periods in Clear and Windy Weather Conditions (2012–2017) Rear-End Crashes on Urban, Divided, Multilane Freeways and Expressways during Peak Periods in Clear and Windy Weather Conditions (2012– 2017) Total Potential Crashes Impacted (2012– 2017) Total Potential Crashes Impacted per Year Blank 28 2 0 2 0.333 PDO 94,526 1,537 286 1,823 303.833 Possible injury 38,608 783 167 950 158.333 Suspected minor injury 9,796 84 10 94 15.667 Fatal and suspected serious injury 892 6 0 6 1 Total 143,850 2,412 463 2,875 479.167 Table B-16. Crashes impacted for short-term timeline. Severity Crashes Reduced per Year (0.5% VMT Share) Crashes Reduced per Year (1% VMT Share) Blank 0.002 0.003 PDO 1.519 3.038 Possible injury 0.792 1.583 Suspected minor injury 0.078 0.157 Fatal and suspected serious injury 0.005 0.010 Total 2.396 4.792 Table B-17. Crash reduction for short-term timeline.

Additional Example Scenarios B-17   Severity Total Crashes on All Roads in All Weather Conditions (2012– 2017) Rear-End Crashes on Urban, Multilane, Divided Roadways during Peak Periods in Clear, Windy, and Rainy Weather Conditions (2012–2017) Rear-End Crashes on Urban, One- Way Roadways during Peak Periods in Clear, Windy, and Rainy Weather Conditions (2012–2017) Rear-End Crashes on Urban, Divided, Multilane Freeways and Expressways during Peak Period in Clear, Windy, and Rainy Weather Conditions (2012–2017) Total Crashes impacted (2012– 2017) Total Crashes Impacted per Year Blank 28 3 0 0 3 1 PDO 94,526 1,772 2 315 2,089 348 Possible injury 38,608 891 2 186 1,079 180 Suspected minor injury 9,796 93 0 10 103 17 Fatal and suspected serious injury 892 7 0 0 7 1 Total 143,850 2,766 4 511 3,281 547 Table B-18. Crashes impacted for medium-term timeline. Severity Crashes Reduced per Year (2% VMT Share) Crashes Reduced per Year (4% VMT Share) Blank 0.010 0.020 PDO 6.963 13.927 Possible injury 3.597 7.193 Suspected minor injury 0.343 0.687 Fatal and suspected serious injury 0.023 0.047 Total 10.937 21.873 Table B-19. Crash reduction for medium-term timeline.

B-18 Framework for Assessing Potential Safety Impacts of Automated Driving Systems The percent crash reduction with respect to average total crashes per year is calculated below using a VMT share of 2%. = − −       × =Percent crash reduction per year 1 23,975 10.94 23,975 100 0.046% The percent crash reduction with respect to average total crashes per year is calculated below using a VMT share of 4%. = − −       × =Percent crash reduction per year 1 23,975 21.87 23,975 100 0.091% Sideswipe Crashes. As previously mentioned, other crash types might be impacted by TJA, such as sideswipe crashes. Sideswipe crashes might increase due to a situation when a vehicle with TJA enabled is following a car in front at a safe distance and an adjacent vehicle sees the gap as an opportunity to switch lanes, causing a sideswipe crash; however, there is a lack of research on this topic. As a point of reference, Table B-20 displays total sideswipe crashes during the study period, 2012 through 2017. This could serve as a point of reference for estimating the relative magnitude of an increase in sideswipe crashes for different assumptions related to percent increase. Interpreting the Results Table B-21 displays a summary of the results for the potential crash reduction of rear-end crashes in comparison to total crashes for different sensitivities of VMT share. The percent crash reduction increases from the short-term to the medium-term timelines as the ODD is expanded to include more facility types and environmental conditions. There is also the potential for TJA to impact sideswipe crashes. However, it is not known exactly how. Table B-20 displays the historical sideswipe crashes in the applicable ODD as Severity Sideswipe Crashes on Urban, Multilane, Divided Roadways during Peak Periods (2012–2017) in Clear Weather Conditions Sideswipe Crashes on Urban, One- Way Roadways during Peak Periods (2012– 2017) in Clear Weather Conditions Sideswipe Crashes on Urban, Divided, Multilane Freeways and Expressways during Peak Periods (2012– 2017) in Clear Weather Conditions Total Crashes Potentially Impacted per Year Blank 0 0 0 0 PDO 740 7 95 140 Possible injury 128 0 16 24 Suspected minor injury 8 0 1 2 Fatal and suspected serious injury 1 0 0 0 Total 877 7 112 166 Table B-20. Sideswipe crashes potentially impacted by TJA.

Additional Example Scenarios B-19   a point of reference. These numbers might increase due to conditional traffic jam assist. The increase in crashes due to TJA is from the potential of other non-ADS-equipped vehicles colliding with the ADS-equipped vehicle. However, there is also potential for sideswipe crashes to decrease that were caused by the ADS-equipped vehicle if they also have a lane-keeping assist feature. Agencies can use the results to estimate the potential reduction in crashes. The following example displays rate reduction for a given agency. An agency is considering infrastructure improvements that will allow use of TJA in additional areas. Based on historical data, there are 30,000 rear-end crashes in a particular area that has 102,000 million VMT. The calculated crash rate is 0.3 rear-end crashes per million VMT or 30 rear-end crashes per 100 million VMT. Using the short- and medium-term VMT shares, the reductions in crash rates and new crash rates are calculated and shown in Table B-22. Assuming a medium-term VMT share of 4%, there will be an expected reduction of 1 crash for every 100 million VMT. An agency could perform a more detailed benefit-cost analysis to compare these benefits with the cost to upgrade or maintain infrastructure as part of the decision process. Communicate Outcomes Through the framework and analysis, the results indicated how the frequency and severity of certain crash types are expected to change. If the ODD is extended so TJA can operate on different facilities or in different environmental conditions, and if data are shared to improve ADS performance, safety will likely change by reducing crash frequency and severity. As the timeline expands and VMT share increases, the percent crash reduction is expected to be greater. Timeline VMT Share Crash Reduction Short-term 0.5% 0.01% 1% 0.02% Medium-term 2% 0.046% 4% 0.091% Table B-21. Summary of results for rear-end crashes. Reduction in Rate (Crashes per 100 million VMT) New Rate (Crashes per 100 million VMT) Short-term: • VMT share = 0.5% 0.2 29.85 • VMT share = 1% 0.3 29.70 Medium-term: • VMT share = 2% 0.6 29.40 • VMT share = 4% 1.2 28.8 Table B-22. Example of crash reduction.

B-20 Framework for Assessing Potential Safety Impacts of Automated Driving Systems However, assumptions were made to perform the analysis. The analysis and results could change slightly as the technology landscape changes, the ODD (or extent of the ODD) expands, and more and better data are available. Two assumptions relate to lane marking and sign conditions. The ODD describes the need for visible lane markings clear of snow or debris and visible signage with high retroreflectivity. That information is not available in the dataset used and not avail- able in typical crash or roadway databases. Because of this, the analysis assumed the lane marking and signage conditions are visible and clear. Another assumption relates to an element in the ODD stating that operating speeds need to be below 35 mph for TJA is operate. The speed of vehicles before a crash is typically not known and not reported on crash records. To create conditions where speeds are less than 35 mph, only crashes occurring during peak periods during the week were used in the analysis. This would hopefully create conditions in which TJA could operate. One last assumption used in the analysis relates to the environmental conditions. The medium- term timeline allows TJA to operate in moderate rain conditions. Crash records do not separate out the level of rain; rather, they only indicate if it was raining or not. The rain variable was used to represent moderate rain conditions. ADS Feature: Highway Platooning (Level 2–Level 4) Understanding the ADS Feature Feature Description In addition to autonomous driving capability scenarios, long- and short-haul freight opera- tions have the opportunity to further increase savings, productivity, and energy efficiency through platooning. Platooning is defined as linking two or more trucks in a convoy using automated driving assistance technologies and with connectivity between vehicles (Figure B-1). This feature can be an advanced cooperative ACC that automates lateral and longitudinal vehicle control of heavy-duty vehicles on a highway. Platooning may include maintaining a formation of vehicles with very short following distances. The feature uses advanced driver assistance system (ADAS) features such as cooperative ACC to maintain short following distances with multiple heavy- duty vehicles. This scenario focuses on highly automated vehicles fitted with second-generation technologies. In addition to providing benefits to the driver, commercial operator, and environ- ment, platooning operations can support capacity management goals. If technology continues to advance, sensor costs fall, and policy adapts, a fruitful system of ADS-equipped trucks deploying platooning features can be achieved. Expected Market A key feature of many scenarios of self-driving truck implementation, platooning is expected to experience market growth. Allied Market Research (2019) valued the global truck platooning Figure B-1. Automated truck and highway platooning scenario (Source: Adopted from Viscelli, 2018).

Additional Example Scenarios B-21   market at $501 million in 2017 and projected it to grow to $4.6 billion by 2025 with a compound annual growth rate (CAGR) of 32.4%. North America currently represents, and will likely con- tinue to represent, the highest share of the global platooning market, though the Latin America, Middle East, and Africa region expects to have the highest CAGR. The study segments the market based on technology, platooning type, communication technology, and region. Techno- logically, the market is classified into groups such as ACC, blind spot warning, global positioning system (GPS), forward collision warning, and lane-keeping assist. Platooning type can be split into driver-assistive truck platooning and ADS-equipped truck platooning. Finally, commu- nication type can be divided into vehicle to vehicle (V2V}, vehicle to infrastructure (V2I), and vehicle to everything (V2X). Policy may be the key driver for the truck platooning market, both in favorable regulation for platoons and new emissions standards that require a decrease in fuel consumption. The production of fully ADS-equipped trucks and the increase in fleet size represent the greatest opportunities for growth in the market. However, the high cost of autonomous technologies and rising security and privacy concerns may limit the speed of adoption and overall growth of truck platooning. Research and pilot studies have been taking place regularly in Europe since 2016. Through regulatory changes and truck manufacturing, the European Union hopes to introduce platooning technologies into the marketplace in 2022. Table B-23 lists a few examples of the ADS-equipped trucks and/or highway platooning models that are commercially deployed or being tested. Anticipated Deployment Scenario Operational Design Domain Table B-24 summarizes the anticipated ODD elements of the ADS-equipped trucks and highway platooning features. The ODD elements are predicted for two timelines, the short-term (next 5 years) and the medium-term (next 5 to 10 years). In addition, the table outlines the major deployment specifications envisioned for the two features that are expected to impact their safety assessment. The deployment elements are also provided for the two timelines. Peloton PlatoonPro Truck Platooning utilizes V2V communications and radar-based active braking systems, combined with vehicle control algorithms. Peloton can link pairs of heavy trucks for connected driving. DAF The EcoTwin uses Wi-Fi-P, radar, and cameras to allow multiple connected trucks to drive close together. Driving inputs are taken from the lead truck, allowing the convoy of vehicles to simultaneously accelerate, brake, and steer. Iveco Iveco participated in the world’s first truck platooning challenge using two trucks from Brussels to Rotterdam. MAN Truck and Bus The company completed a successful pilot where two professional drivers drove electronically linked trucks on the Autobahn for 35,000 test kilometers over 7 months. Volvo Volvo, FedEx, and the North Carolina Turnpike Authority used ADAS to test the first public highway platooning showcase in the United States using three Volvo VNL tractors. Tesla The Semi is an electric Class 8 semi-trailer truck with a stated goal to provide auto pilot and based platooning capabilities using radar and cameras led by a truck with a driver. Table B-23. Highway platooning examples.

B-22 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Operational Design Domain Level Timeline Additional Deployment Context Self-driving features of ADS-equipped trucks can be supplemented with platooning capabilities to further increase benefits. • Freeways (both urban and rural). • Dedicated lanes. • Navigate through interchanges and ramps. Short-term (high disruption). • The lead truck in a platoon has a driver who must take control as needed. Following trucks use the platooning feature to follow the leader. • Platoons of two to three trucks. • V2V communication. • Following trucks are Level 3; lead trucks can be Level 2 or Level 3. • Operating on well-marked roads with well-maintained signs. • Freeways (both urban and rural). • All +4-lane divided highways (urban and rural). • Mixed traffic. • Navigate through interchanges and ramps. Medium- term (high disruption). • The lead truck and following trucks can be driverless. • Platoons of three or more trucks. • V2X communications. • All trucks in platoon are Level 4. • Operating on well-marked roads with well-maintained signs. Table B-24. Highway platooning deployment scenarios of interest. Table B-25 provides a summary of the technology specifications and key infrastructure needs pertinent to the envisioned highway platooning scenarios. Stage of Technology Development • First Generation (Gen I): This is the first version of ADS-equipped trucks sensor package and their underlying computational algorithms for processing the data. Typically, this pack- age embraces the needed combination of sensors such as forward-facing cameras, radar, ultrasonic sensors, laser scanners, and inertial measurement units (gyroscopes and acceler- ometers) with a priori digital maps (lane-level detail). This package will not have a good object detection capability in low-visibility conditions, limiting the ODD to certain conditions (e.g., light rain, no snow, good lane markings). Similarly, the underlying perception algorithms for processing data are capable of hand ling the basic computations for the proper functionality of the ADS-equipped truck feature within the ODD and deployment context specified in Tables B-24 and B-25. These algorithms are at early development and may have more errors than later, more mature technology, leading to lower safety performance and/or lower percentage of time operating in autonomous mode (high disengagement rate). Sensors and computation algorithms used at this stage are com- mercially available and currently operate in certain vehicles. • Second Generation (Gen II): Gen II is a more advanced stage of technology than first-genera- tion models or systems. In addition to the sensor types included in Gen II, this generation embraces high-fidelity lidar sensors and an OBU. All Gen II sensors are newer, more advanced, accurate, and have longer perception range than the Gen I sensors. A key feature of the Gen II sensor package is providing layers of redundancy to one another. The percep- tion algorithms frequently cross-check the data from different sensors to ensure that no object is left undetected and to eliminate false positives.

Additional Example Scenarios B-23   • This sensor package manifests as mature sensor fusion technology that is able to combine the sensing capabilities of multiple sensors, resulting in more reliable and robust per- ception with a broad sensing scope. To this end, the underlying perception algorithms for processing the data are more advanced and are capable of performing complex sensor fusion calculations enabling the operation in an expanded ODD and deployment con- text. Another envisioned key feature of this generation of sensors is integrating V2V and V2I communication within the vehicles through the OBU. This provides opportunities for the vehicle to receive real-time dynamic data for weather, work zones, and traffic. The new advanced sensor suite allows trucks to operate effectively on more road types, such as four- lane divided highways, in more severe weather, and on roadways with imperfect lane mark- ings and signage. Table B-26 provides a qualitative assessment of the technology state of the different sensor package generations. The table highlights the key functional and technical differences between the two generations as well as the operational atmospheric conditions. Infrastructure Needs and Impacts The features assist the driver navigating a highway, and when in platooning mode, the follower vehicle maintains a headway and lateral alignment. The sensors on board the vehicle need to detect infrastructure elements, such as lane markings, barriers, and signs, to determine proper heading and speed. The infrastructure requirements for this feature are largely driven by chal- lenges in human factors, connectivity, and limits to ADS perception technology. To increase functionality and efficiency of autonomous and platooning truck features, ADS-equipped truck ports (ATPs) may need to be constructed near interstates. At ATPs, drivers Expected Timeline Vehicle Type Sensor Package Key Infrastructure Requirements Digital Physical Short-term High disruption Heavy- duty. Gen I. • V2V communications. • GPS. • HD maps. • Weather data. • Infrastructure data. • Clear lane markings. • Visible signage. • Highly detectable traffic control device. • Work zone alerts. Medium-term High disruption Heavy- duty. Gen II. • V2X communication. • GPS. • HD maps. • Weather data. • Infrastructure data. • Work zone alerts. • 5G and dedicated short- range communication. • Lane markings. • Visible signage. • Highly detectable traffic control device. Table B-25. Highway platooning key infrastructure requirements.

B-24 Framework for Assessing Potential Safety Impacts of Automated Driving Systems operating locally can swap trailers to autonomous tractors optimized for highway driving. Likewise, highway-optimized trucks can swap trailers to human-driven trucks for last-mile and urban delivery where driverless operations are more complex. In platooning, the leader vehicle’s responsibilities differ from those of the follower vehicle. For the rear vehicle, the radar and camera are critical for measuring distances, and connectivity provides information about the lead vehicle’s future speed, which is critical for maintaining the headway and lateral position. Look-ahead information about traffic disruptions, such as congestion or emergency vehicles, can help improve coordination between vehicles in a platoon. Connectivity can provide additional information that improves freight operations and safety in targeted areas. For example, along Wyoming roadways over 6,000 ft in elevation, wind gusts exceed 65 mph and crash rates are three to five times as high in winter than in summer. The Wyoming DOT pilot is developing applications that use V2I and V2V connectivity to support a range of services from advisories including roadside alerts, parking notifications, and dynamic travel guidance to help reduce the number of blow-over and adverse weather incidents in the corridor to improve safety and reduce incident-related delays. Partially automated platooning features could benefit from these types of applications. Risk Assessment Risks • A major risk is navigating the “machine-to-human handover,” when the machine driver requests to hand back control to the human. Since it is irresponsible for the machine driver to simply signal to the human “Here, you take over,” it is evident there must be a period of time follow- ing the handover request for the human driver to regain proper situational awareness. • At present, ADS and platooning features are variable in trucks. With no standard on how the ADS feature performs, which sensors it includes, or in what circumstances they can platoon, Scenario Scenario Technology State Comparison Qualitative Assessment of Technology State Operational and Atmospheric Conditions Key Functional and Technical Differences • Gen I • Smaller platoons. • ODD exclusively on freeways. • Communications with other trucks. Weather: clear, wind. • V2V communications. • Two- to three-vehicle platoons. • Gen II • Larger platoons. • ODD expanded to divided highways. • Communications with vehicles and roadway infrastructure. Weather: clear, wind, rain. • Lidar. • V2X communications. • Platoons of three or more vehicles. Table B-26. Scenario technological specifications of highway platooning.

Additional Example Scenarios B-25   a driver may not be familiar with more than one machine-user interface and may fail to react safely in a request for control takeover initiated by the ADS. • The occupant injury risks in truck platooning are different from risks in single-passenger vehicle crashes. Structural adequacy, vehicle stability, and occupant risk are the main criteria when evaluating roadside infrastructures. • ADS-equipped truck features and platoons may contribute to new crash types. Though rear- end crashes may become less frequent, large trucks moving closely together may promote different driving behavior when on mixed roads, increasing human-driven lane change risk for drivers next to platoons. • Platoon length, speed, and intraplatoon gaps may vary over trips. In real traffic conditions, fuel reductions may not be observed because speeds and convoy gaps will vary due to inter- actions and disruptions with other vehicles (Ramezani et al., 2018). • Daimler, the German automotive corporation, recently said it would “reassess its view on platooning.” Daimler tested platooning for several years in the United States and deter- mined that benefits were “less than expected.” Even in perfect driving conditions, fuel savings were low and diminished further when the convoy becomes disconnected and trucks must accelerate to reconnect. Because of this, Daimler determined that in U.S. long-distance appli- cations, analysis showed no business case (Lopez, 2019). • Relatively low numbers of units are sold by truck manufacturers (Viscelli, 2018). • Roadway infrastructure, such as bridges, may not be able to handle the higher weight concen- tration that platoons create (Grant, 2018). • Lateral wandering of AV platoons is much narrower than that of human-driven trucks, increasing pavement cracking and rutting. Pavement fatigue in turn increases the risk of hydroplaning (Zhou et al., 2019). • Labor may be opposed due to job loss. • Training is needed on ADS and ATP use. Opportunities • Truck platooning is a promising technology that could bring great benefits to society and road users. In fact, the wide benefits achieved by self-driving trucks and highway platooning (e.g., safety, energy savings, road capacity) are expected to be the main reason for expanding the market of this feature more rapidly than other features for passenger cars. • Unlike cars, there is already high demand for ADS-equipped trucks. Because of the labor savings of autonomy and because trucks are bought as business decisions thoroughly evalu- ated by fleets, return on investment on ADS-equipped trucks is expected to be high (Viscelli, 2018). In 2013, Moran Stanley estimated that ADS-equipped trucks would provide $168 billion in savings (Viscelli, 2018). • Implementing ATPs will provide a host of benefits to both industry and drivers at risk from automation. ATPs can be built in strategic locations near interstate exits and truck parking lots outside congested urban areas. ATPs not only allow for trailer switching, but also provide driver facilities and refueling and charging stations. ATPs could facilitate off-peak deliveries to reduce road congestion and cut down on the coordination between shippers and carriers. Efficiency could be greatly enhanced through a ridesharing-style service that matches drivers and freight through an app with real-time pricing, keeping wages and work opportunity high (Viscelli, 2018). • Platooning is expected to reduce the air resistance within a traveling fleet, which can trans- late to significant savings in fuel consumption by the fleet. The added advantage of reducing aerodynamic drag can reduce the maximum fuel consumption of trucks by between 6.5% and 21% (Ramezani et al., 2018) and induce fuel reductions of entire fleets up to 43% (Eilbert et al., 2019).

B-26 Framework for Assessing Potential Safety Impacts of Automated Driving Systems • In the future, many trucks with ADS and platooning capabilities will likely be electric. In the United States, the transportation sector is responsible for almost 30% of annual greenhouse gas (GHG) emissions (US EPA, 2019). Battery electric vehicles (BEV), however, have been shown to reduce overall GHG emissions and pollution relative to vehicles with an internal combustion engine and could greatly reduce our need for oil (Notter et al., 2010; Delucchi et al., 2014). Transferring emissions from the tailpipe to power-generating plants also further centralizes total emissions in the power production sector where measures such as carbon capture and sequestration and a cleaner fuel mix could contribute to reduced overall emissions. Additionally, BEV engines are inherently more energy efficient than internal combustion engines (Messagie, 2017) and can increase energy efficiency further by making use of “energy recovery” technology where breaking and unaccelerated motion act to recharge the battery (Delucchi et al., 2014; Manzetti and Mariasiu, 2015). Also, this feature eliminates the need for a highly skilled driver in following trucks, which could bring shipping costs down when platooning is deployed at a large scale. • ADS-equipped trucks and platooning could positively affect other road users by providing increased capacity for the road (reducing gaps between vehicles). Self-driving and platooning also offer safety benefits when applied at a large scale by eliminating truck drivers’ induced errors. • Roadside barriers have proven to produce low risk of injury to platooning truck occupants given failure. Goals and Hypothesis For the deployment scenario of highway truck platooning, the goal is to reduce the frequency and severity of truck-involved crashes on truck-route corridors through use of highway truck platooning (SAE Level 4) and supporting infrastructure. The overall hypothesis is that highway truck platooning will improve safety on interstates, freeways, and other divided highways with four or more lanes by reducing truck-involved crashes. The expected change in the number and percentage of truck-involved crashes will depend on market penetration, which is explored in the analysis. The questions to evaluate this hypothesis are listed below and relate to crash types, crash severity levels, infrastructure, and data. 1. How will the frequency of crashes change? It is anticipated that highway truck platooning will change the frequency of all crashes. While it is anticipated that truck-related or truck- involved crashes would decrease, other types of crashes might increase, such as crashes in the vicinity of interchanges where merging or diverging maneuvers occur. 2. How will the severity of truck-involved crashes change? It is anticipated that the severity of truck-related and truck-involved crashes would decrease. It is also anticipated that the severity of other crashes would remain the same as the speed, angle of impact, and other factors related to severity are expected to remain the same. 3. How will the number of trucks on interstates and highways change and, in turn, impact the number of crashes? While overall it is anticipated that highway truck platooning would reduce crash frequency and severity by removing human error, it is also anticipated that truck volume will increase with the deployment of highway truck platooning due to trailing trucks not needing humans present in the trucks. This could alter the crash frequency due to the increase in truck traffic. 4. Will the safety and deployment of highway truck platooning change if the ODD is extended in which highway truck platooning can operate? The current ODD for the short- and medium- term timelines includes certain weather conditions and certain infrastructure conditions. If the ODD is extended, it is anticipated that more crashes could be impacted by the deploy ment of highway truck platooning.

Additional Example Scenarios B-27   In summary, through the deployment of highway truck platooning, it is hypothesized that the frequency and severity of truck-involved and truck-related crashes will decrease. While other crash types such as merging and diverging actions at interchanges may increase, it is expected that the severity distribution will remain the same. While this provides an overview of potential safety impacts, it is important to perform a crash sequencing exercise to think through the contributing factors and precipitating events that lead to a crash. The following are a few examples related to truck platooning along the interstate: 1. Run-off-road: a. Driver of truck is distracted, falls asleep, or is otherwise inattentive and the vehicle drifts off the road. b. Driver of truck is fully attentive and adverse weather contributes to the driver losing control or incorrectly navigating and the vehicle leaves the road. c. Driver of truck is fully attentive and sudden congestion leads to an evasive maneuver whereby the driver attempts to avoid the back of queue and the vehicle leaves the road. 2. Rear-end: a. Driver of truck is distracted, falls asleep, or is otherwise inattentive and the truck rear-ends another vehicle. b. Driver of truck is fully attentive and adverse weather contributes to limited stopping dis- tance whereby the driver is not able to stop or slow and the truck rear-ends another vehicle. c. Driver of truck is fully attentive and sudden congestion leads to unanticipated braking whereby the driver attempts to stop, but the truck rear-ends another vehicle. From the anticipated capabilities of ADS-supported truck platooning and the above crash sequencing, the project team identified specific opportunities for this ADS application to mitigate crashes. For example, ADS-supported truck platooning is not expected to operate in adverse conditions, so there is limited potential to mitigate crashes related to sequence 1b; however, ADS-supported truck platooning is expected to provide opportunities to mitigate crashes related to sequence 1a and 1c. Similarly, ADS-supported truck platooning is not expected to mitigate crashes related to sequence 2b but is expected to mitigate crashes related to sequence 2a and 2c. Methods of Analysis Components of the Analysis. To estimate the safety effects of highway truck platooning, an analysis is needed to quantify the potential impacts by using crash data and the types of crashes platooning is expected to influence. The ODD of highway truck platooning is separated by facilities and traffic features in which they can operate in short- and medium-term timelines. For both timelines, highway truck platooning can operate on interstates and freeways and can navigate through interchanges and ramps. However, highway truck platooning can operate only in dedicated, separated trucking lanes for the short-term timeline but can operate in traditional lanes with mixed traffic during the medium-term timeline. Because of the lack of dedicated truck lanes in transportation networks and lack of data, dedicated lanes cannot be filtered out, and it is assumed that the ODD for the short-term timeline includes interstates and freeways. Highway truck platooning can also operate on urban and rural divided highways with four or more lanes, in addition to interstates and freeways, in the medium-term timeline. Highway truck platooning is anticipated to only operate during certain weather conditions. In the short term, it is anticipated that platooning will operate in clear and windy weather con- ditions (e.g., no rail, snow, ice, etc.). In the medium term, it is anticipated that platooning will operate in clear, windy, and rain weather conditions. Truck-involved crashes are crashes where at least one vehicle involved, per the police record, is a truck. While truck-involved crashes are the main crash type impacted, there is the potential to impact crashes with no trucks indicated on police records [e.g., truck-related action contributed

B-28 Framework for Assessing Potential Safety Impacts of Automated Driving Systems to crashes involving other vehicle(s) on the road], which are referred to as truck-related crashes. However, in this analysis, only truck-involved crashes are assessed because of the lack of infor- mation from police reports to determine whether crashes are truck-related. It is anticipated that all crash severity levels will be impacted by highway truck platooning. A summary of the components of the analysis is shown in Table B-27. It is assumed that highway truck platooning will not eliminate all truck-involved crashes in the ODD described above. The assumption is that platooning will eliminate only a certain percentage of crashes because not all trucks will be equipped with the ability to platoon, and if a truck is able to platoon, there is the possibility that it is not used or does not operate as intended. It is assumed that highway truck platooning will also have an impact on merging and diverging actions in the vicinity of ramps and interchanges, potentially causing an increase in crash fre- quency. The analysis related to this identifies whether a vehicle maneuver was either “merge” or “diverge” but does not show a potential crash reduction due to lack of information on exactly how these crash types will be impacted. The expected timeline for achieving success depends on the expected timeline for deployment and penetration rates of platooning trucks. In this scenario, the timeline is assumed to pertain to short- and medium-term planning, which is a high disruption. Analysis Methods Data Sources. The sources of crash, roadway, and traffic data to analyze the impact of highway truck platooning are similar to those of ADS-equipped trucks due to similar ODDs, distribution, and VMT share. The data from HSIS include crash, roadway, and traffic data from Illinois and Ohio. The data were merged to identify crashes that could be impacted by highway truck platooning. Illinois data are from 2006 to 2010, and the Ohio data are from 2010 to 2015. These merged data were used in the analysis described below. Define Metrics. The metrics are also similar to those of ADS-equipped trucks due to similar ODDs, disruption, and VMT share. The primary metric for assessing the safety impact of highway truck platooning is the change in crash frequency and severity. Merge and diverge vehicle maneuvers in vicinity of interchanges. Crash severity level. All severity levels. Traffic conditions. Dedicated lanes (short-term). Mixed (medium-term). Component Description Functional classification, lanes, and road configuration. Interstates and freeways (short- and medium-term). All other functional classifications that are divided with four or more lanes (medium-term). Area type. Urban and rural. Speed conditions. Any speed limit. Weather conditions. Clear, wind (short-term). Clear, wind, rain (medium-term). Crash types. Truck-involved. Table B-27. Components of the analysis.

Additional Example Scenarios B-29   Evaluation Method. The impact method is also similar to that of ADS-equipped trucks. The evaluation uses crash history and identifies the crashes that could potentially be affected by the use of ADS-equipped trucks. The evaluation also includes a sensitivity analysis to explore assumptions related to penetration rates as well as probabilities that trucks are platooning and that the feature is activated and functions properly. Results Crash Reduction Due to Highway Truck Platooning. The impact of highway truck pla- tooning is similar to that of ADS-equipped trucks due to their similar ODDs, distribution, and VMT share. However, the ADS-equipped trucks analysis does not include the potential crash increase due to merging and diverging. Based on the ADS-equipped truck analysis, Table B-28 displays the percent reduction in total truck-involved crashes per year for various sensitivities of VMT share in the short- and medium-term timelines. Crash Increase Due to Highway Truck Platooning. While some crashes are expected to decrease, as shown previously, there is potential for other types of crashes to increase. Crashes related to merging and diverging maneuvers at ramps can potentially increase. Vehicles merging or diverging from a ramp to the main line might have difficulty merging into or out of traffic if a truck platoon is driving past the ramp or interchange. The following sections discuss crashes that could increase due to highway truck platooning; however, no analysis was performed due to the lack of information on how highway truck platooning will impact vehicles merging or diverging to or from ramps and the mainline. Illinois data were used to identify the number of crashes involving merging or diverging maneuvers. The Illinois data were selected because of the availability of crash codes for vehicle maneuver and the explicit indication of a vehicle merging or diverging. Table B-27 shows the ODD (same as for the ADS-equipped truck analysis), indicating the filters for roadway types and weather conditions for both the short- and medium-term timelines. Table B-29 displays average total crashes per year for the short-term timeline for all crashes and truck-involved crashes by severity level when a vehicle maneuver in the crash was “merge” or “diverge.” Table B-30 displays average total crashes per year for the medium-term timeline for all crashes and truck-involved crashes by severity level when a vehicle maneuver in the crash was “merge” or “diverge.” While the average crashes per year where vehicle maneuvers were “merge” or “diverge,” it is unknown how highway truck platooning will affect these numbers. It is anticipated that crashes involving merging or diverging maneuvers could increase because of highway truck platooning. Interpreting the Results Agencies can use the results to estimate the potential reduction in crashes. The following example displays rate reduction for a given agency. The agency is considering converting its truck Timeline VMT Share Percent Reduction in Total Truck-Involved Crashes per Year Short term 5% 0.135% 10% 0.271% Medium term 11% 0.471% 22% 0.941% Table B-28. Components of the analysis.

B-30 Framework for Assessing Potential Safety Impacts of Automated Driving Systems fleet to be automated and able to drive in platoons or improving infrastructure to allow highway truck platooning (e.g., constructing truck-only lanes or truck-bypass lanes). Based on historical data, there are 5,000 heavy-truck-involved crashes in an area with 2,000 million heavy-truck miles traveled. The calculated crash rate is 250 crashes per 100 million heavy-truck miles traveled. Using the short-term and medium-term VMT share, the calculated reduction in crash rates is in Table B-31. Assuming a short-term VMT share of 5%, a reduction of 12.5 crashes is expected for every 100 million heavy-truck miles traveled. An agency could perform a simple or more detailed benefit-cost analysis to compare these benefits with the cost of converting a truck fleet or the cost of infrastructure improvements or maintenance as part of the decision process. Communicate Outcomes The analysis supports the goals and hypotheses of the safety impacts of highway truck pla- tooning. The results determined that truck-involved crashes would decrease in frequency and severity with the deployment of highway truck platooning. There is the possibility that crashes at interchanges and ramps would change; however, that could not be verified. Severity Level Total Crashes per Year in Clear and Windy Weather Conditions with Vehicle Maneuver of Merge or Diverge on Interstates and Freeways Truck-Involved Crashes per Year in Clear and Windy Weather Conditions with Vehicle Maneuver of Merge or Diverge on Interstates and Freeways PDO 530 285 Possible injury 18 7 Suspected minor injury 33 15 Fatal and suspected serious injury 11 3 Total 591 310 Table B-29. Crashes with merging or diverging vehicles for short-term timeline. Severity Level Total Crashes per Year in Clear, Windy, and Rain Weather Conditions with Vehicle Maneuver of Merge or Diverge on Interstates, Freeways, and Divided Roads with 4 or More Lanes Truck-Involved Crashes per Year in Clear, Windy, and Rain Weather Conditions with Vehicle Maneuver of Merge or Diverge on Interstates, Freeways, and Divided Roads with 4 or More Lanes PDO 971 385 Possible injury 47 12 Suspected minor injury 61 20 Fatal and suspected serious injury 21 5 Total 1,100 422 Table B-30. Crashes with merging or diverging vehicles for medium-term timeline.

Additional Example Scenarios B-31   Assumptions were made regarding facility conditions. Roads need to be well marked and well maintained for highway truck platooning to operate. This information is not readily available in the data used and is not available in most roadway databases. Therefore, it had to be assumed that the roadways were well marked and well maintained. The ODDs of ADS-equipped trucks and highway truck platooning were similar. Therefore, it was assumed that truck-involved crash reduction would be the same. Crashes around inter- changes and ramps could be different from those of ADS-equipped trucks, but that could not be determined with the available data and information about highway truck platooning. ADS Feature: Level 4 Fleet-Operated Automated Driving System–Dedicated Vehicle (ADS-DV) Understanding the ADS Feature Feature Description. Level 4 fleet-operated ADS-DVs are driverless vehicles anticipated to be operated primarily by transportation network companies. This ADS feature allows the vehicle to operate without the need for a driver on a predefined set of roads or geographic area and within the specified ODD. The ADS automatically collects and processes data from onboard sensors and handles the V2V communications to perceive the surroundings, as well as obstacles and relevant signage, and identify the appropriate navigation trajectories. Fleet-operated ADS-DVs are capable of navigating through the identified trajectories safely with no human input. For this feature, the ADS is responsible for the DDT fallbacks and for achieving a minimal risk condition. Fleet-operated ADS-DVs that are designed to also accommodate operation by a driver (whether conventional or remote) may allow a user to perform the DDT fallback if they choose to do so. However, a Level 4 need not be designed to allow a user to perform DDT fallback and, indeed, may be designed to disallow (SAE, 2018). For instance, consider a scenario where a fleet-operated ADS-DV, which handles the entire DDT within a geo-fenced area, experiences severe weather conditions that are beyond its ODD. In response, the fleet-operated ADS-DV should handle the DDT fallback and achieve a minimal risk condition that could include turning on the hazard flashers, maneuvering the vehicle to the road shoulder and parking it, and automatically summoning emergency assistance, which concludes the fallback response. However, if a driver was allowed to handle the DDT fallback and accepted risk, the ADS-DV may simply continue driving manually instead of achieving a minimal risk condition. For noticeable benefits from Level 4 fleet-operated ADS-DVs, it is imperative that they operate as electric vehicles for many reasons. First, vehicle miles traveled are expected to increase drastically Reduction in Rate (Crashes per 100 Million Heavy-Truck Miles Traveled) New Rate (Crashes per 100 Million Heavy-Truck Miles Traveled) Short term: • VMT share = 5% 12.5 237.5 • VMT share = 10% 25.0 225.0 Medium term: • VMT share = 11% 27.5 222.5 • VMT share = 22% 55.0 195.0 Table B-31. Example of crash reduction.

B-32 Framework for Assessing Potential Safety Impacts of Automated Driving Systems for this level of automation if they are operated as shared vehicles. If fleet-operated ADS-DVs are offered as gasoline engines, their operational costs will outweigh the benefits of shared mobility. Additionally, the environmental impacts would make them an unpopular option and, for AVs to operate safely, vehicle sensors and components must be maintained in good condition and operate with high levels of reliability. This is difficult to maintain in an internal fuel combustion environ- ment, where there is a need to maintain all components of the engine in a reliable manner. Overall, it is much easier and more reliable to maintain the fleet-operated ADS-DV sensors and components functioning well in a drive-by-wire environment compared to an internal combustion engine. OEMs are aware of these facts, and to facilitate the commoditization of their offered products, they are offering them as electric vehicles. Major cities are concerned with the potentially nega- tive implications of increased shared gasoline vehicles and have started scheduling a future ban for fuel combustion engines (Boffey, 2019; Evarts, 2019). This trend is expected to intensify with a successful commercial deployment of Level 4 fleet-operated ADS-DVs, assuming there is public acceptance to support this engagement. Expected Market. In December 2018, Waymo officially launched its commercial fleet- operated ADS-DV service in the suburbs of Phoenix (Hawkins, 2019b). During early deploy- ment stages of fleet-operated ADS-DV, the car manufacturers will primarily target transportation network companies (TNCs) with large fleets. Large-fleet companies will be able to afford the HAV at early deployment and will be the key players for expanding the HAV market. In an early response to these facts, automakers and TNCs have been forging corporate coalitions to reserve a decent share of what is expected to be a multibillion-dollar market by 2030. For instance, in 2017, Volvo announced a 3-year deal to supply up to 24,000 ADS-DV to Uber, and the vehicle ship- ment was expected to start by the end of 2019. On the other hand, Lyft began partnering with General Motors earlier in 2016 for development of HAVs. Lyft pursued further collaboration by making two separate pacts in 2017 with Ford and Waymo, respectively, for collaboration on fleet-operated ADS-DVs. Accordingly, the market of TNCs and shared mobility is expected to grow rapidly with the introduction of fleet-operated ADS-DVs reaching $173.15 billion by 2030, with shared mobility services contributing to 65.31% (Frost and Sullivan, 2018). These cascade trends and partner- ships are expected to reshape the transportation sector and trigger a disruptive change to the transportation industry, probably the largest in transportation history. For instance, by 2030 the traditional processes for buying and renting cars could be replaced by a subscription system operated by TNCs and fleet managers. These systems would offer a variety of ride-hailing plans covering the different types of subscriber trips (e.g., work, errands, recreational). Discounts would be offered for subscribers affiliated with the TNC on long-term agreements to a level that makes the cost per mile traveled using these services much cheaper and more convenient than the cost and satisfaction per mile traveled using private cars. Assuming these scenarios, parking lots are expected to shrink across the United States and gradually be replaced with electric charging stations. Cities are partnering on tests of a variety of fleet-operated ADS-DV technologies, including retrofitted autos, and shuttles and innovative types of vehicles like conveyors (small, cart-sized AVs that travel on sidewalks). The ODD for testing of these pilots can take place on private roads and planned environments, such as technology parks and college campuses, or public roads and city streets. AV pilots take many forms but can be categorized into two main groups: (a) HAVs for TNCs and (b) low-speed shuttles. The next subsections review some of the pilots conducted and ongoing and outline the lessons learned from some of these pilots (Table B-32). Anticipated Deployment Scenario Operational Design Domain. This scenario assumes that fleet-operated ADS-DV deploy- ments will initially limit coverage only to certain trips within a geographic boundary. The ODD is

Additional Example Scenarios B-33   assumed to be geographically bounded, and the designated route network for the fleet-operated ADS-DV does not use all roads and intersections within the road network (e.g., highways, school zones). As an illustrative example of an early deployment geographic region, consider Chandler, Arizona. This ODD includes signalized intersections, stop signs, pedestrian crosswalks, bike lanes, multilane roads, and speed limits up to 45 mph. It does not include some roads, including those over 45 mph and temporary traffic zones. The infrastructure in this ODD is well suited to early deployment, including modern and well-maintained signage and markings, good sight distance, minimal grade and curvature, and wide lanes. Table B-33 summarizes the anticipated ODD elements of the fleet-operated ADS-DV fea- ture. The ODD elements are predicted for two timelines: the short term (next 5 years) and the medium term (next 5 to 10 years). In addition, the table outlines the major deployment specifi- cations envisioned for the two features that are expected to impact their safety assessment. The deployment elements are also provided for the two timelines. Uber Pittsburgh, PA; Tempe, AZ (ended) In September 2016, Uber began a SAV pilot in Pittsburgh, and was the first SAV service in the United States to serve passengers selected from the public. However, testing stopped in both cities after the high-profile crash and death in Tempe, Arizona, in 2018. While Uber is banned from testing in Arizona, it began testing SAVs in “manual mode” with a specialist in control at all times in Pittsburgh (Bliss, 2018; Rogers, 2018). Voyage The Villages, FL Voyage operates SAV pilots at The Villages retirement community in Central Florida. Service launched in Florida in 2018 (Corder, 2018). Ford/ Domino’s Ann Arbor, MI; Miami, FL A Ford Fusion hybrid began delivering pizzas with test driver in Ann Arbor in 2018 and plans were announced for Miami (Marakby, 2018). Voyage The Villages, San Jose, CA Voyage operates SAV pilots at The Villages retirement community in San Jose. It has operated in San Jose since 2017 (Cameron, 2017). Operators Location Description Waymo Phoenix, AZ Waymo launched an Early Rider program in early 2017, allowing select Phoenix residents to request rides in their automated minivans. Waymo engineers have now moved to the backseat as of November 2017 ( Barr, 2018). Cruise/GM San Francisco, CA In 2017, Cruise launched its pilot, “Cruise Anywhere,” a shared AV (SAV) service for its employees to use for preselected destinations in San Francisco. Cruise intended to launch a commercial SAV offering in 2019 (Felton, 2018). Nuro and Kroger Foods Scottsdale, AZ As of August 2018, Nuro is running its grocery delivery pilot using Toyota Priuses and intended to start delivery with its specialized R1 vehicle in fall 2018. The R1 is designed to exclusively have space for delivery goods, without any passengers (Nuro, 2018). NuTonomy and Aptiv Boston, MA NuTonomy has been testing its vehicles in the Seaport neighborhood of Boston since 2017, and in June 2018, the vehicles were approved for testing citywide. NuTonomy is required to submit quarterly update reports to the city of Boston (Locklear, 2018). Table B-32. Recent AV pilot programs.

B-34 Framework for Assessing Potential Safety Impacts of Automated Driving Systems • Cannot navigate temporary traffic zones (e.g., work zones, school zones). • Sidewalks rarely traversed by pedestrians. • Limited VRU interactions, mostly traveling within designated areas (e.g., crosswalks and bike lanes). • Can operate in any sky condition, 3 with the exception of conditions that cause sun glare.4 Operational Design Domain Level Timeline Additional Deployment Context • Suburban geofenced area. • All streets and intersections within a typical suburb such as Chandler, Arizona, that are fulfilling an ODD model set by the manufacturer. o Operating on well- marked roads with well- maintained signs. o Operating only in clear and good weather condition (e.g., no rain, snow, etc.). o Speed can reach up to 45 mph. Short term • Market share = 10% • Fleet share = 7% • VHT share = 10% • Level 4 automation ADS feature where ADS is responsible for the DDT fallback and achieving appropriate minimal risk conditions. • Operating predominantly as fleet- operated ADS-DVs operated by transportation network companies. • Increased willingness for vehicle sharing. • Decreased vehicle ownership. • Navigating among other road users, including passenger vehicles, commercial vehicles, transit vehicles, and vulnerable road users. • Navigating among special vehicles, such as emergency vehicles, school buses, and mail delivery vehicles. • Highly visible signalized intersections with backer plates. • Road surface conditions in very good condition (Pavement Condition Index, 2019)1 with minimal road damage (e.g., cracking, rutting, raveling, potholes). • Road surface is dry and not covered by snow or ice. • Road markings are in good condition.2 Table B-33. Fleet-operated ADS-DV deployment scenarios of interest. 1 Very good state of repair is defined as 80–89.9 Pavement Condition Index (PCI). 2 Refer to NCHRP Project 20-102(06) for more details on road markings for machine vision (Pike et al., 2018). 3 The National Weather Service provides a scale of sky conditions by percent of opaque cloud cover. 4 Sun glare occurs if the sun is within a specified angular distance between the driver’s line of sight and the sun. For a 40-year-old driver, this angular distance is 19° and for a 60-year-old driver it is 25°. For early deployments, 25° is assumed. This would exclude travel in the direction of the sunrise and sunset while the sun is below this angle.

Additional Example Scenarios B-35   • Suburban and urban geofenced area. • All streets and intersections within a suburb or a central business district that are fulfilling an ODD model set by the car manufacturer. o Operating on well- marked roads with well- maintained signs. Medium term • Market share = 30% • Fleet share = 14% • VHT share = 22% • Level 4 automation ADS feature where ADS is responsible for the DDT fallback and achieving appropriate minimal risk conditions. • Operating predominantly as fleet- operated ADS-DVs operated by TNC. • Increased willingness for vehicle sharing. Operating in good and clear weather conditions but could operate in mild and light rain conditions. o Speed could reach up to 50 mph. • Decreased vehicle ownership. • Sidewalks are frequently traversed by pedestrians. • Frequent pedestrian interaction with the vehicle. • Up to moderate rain, but no snow, sleet, freezing rain, or hail.7 o • Illuminance is brighter than twilight (AMS Glossary of Meteorology, 2018; NOAO, 2019).5 • No rain, snow, sleet, freezing rain, hail, or more than gale force winds (NOAA, 2019). 6 Operational Design Domain Level Timeline Additional Deployment Context Table B-33. (Continued). 5 Twilight may be defined multiple ways. The National Optical Astronomy Observatory has developed a scale for illuminance, or intensity of visible light, which is measured in Lux. Twilight is illuminance greater than 10.8 Lux. The American Meteo- rological Society describes civil twilight as when the sun is 6 degrees below the horizon, http://glossary.ametsoc.org/wiki. 6 Beaufort Wind Scale, including World Meteorological Organization Classification. 7 Moderate rain is defined as up to 0.3 inch per hour (0.76 cm/hr), according to the American Meteorological Society, Glossary of Meteorology, 2019, http://glossary.ametsoc.org/wiki. Table B-34 provides a summary of the technology specifications and key infrastructure needs pertinent to the envisioned fleet-operated ADS-DV scenarios. First Generation (Gen I): This is the first version of sensors and technology package and the underlying computational algorithms for processing the data needed for full autonomy. Typically, this package embraces all needed combinations of sensors such as forward- facing cameras, radars, ultrasonic sensors, lidars, and inertial measurement units (gyro- scopes and accelerometers) with a priori HD maps. This package will not have a good object detection capability in low-visibility conditions, limiting the ODD to certain conditions (e.g., good and clear weather condition, good lane markings). Similarly, the underlying perception algorithms for processing the data are at early stages and are capable of handling the basic computations for reaching full autonomy within the ODD and deployment context specified in Table B-34. These algorithms are in early development and may have more errors than later, more mature technology, leading to lower percentage of time operating in autonomous mode (high disengagement rate). Sensors and computation algorithms used at this stage are commercially available and currently operate in certain vehicles (Hawkins, 2019a).

B-36 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Second Generation (Gen II): Gen II is a more advanced stage of technology than first-generation models or systems. In addition to the sensor types included in Gen I, this generation embraces a vehicle OBU for handling V2X communications. All Gen II sensors are newer, more advanced, more accurate, and have longer perception range than the Gen I sensors. Assuming a high-disruption scenario would occur, the development curve for these technologies could follow Moore’s Law in computer chips. For instance, it is possible that every 18 months, resolution doubles and the price drops by half for the lidar sensor—the most important and expensive sensor in the package (Bloomberg, 2019). The supporting HD maps for this generation have better accuracy and include more reliable dynamic layers—updating in real time—for different traffic conditions, such as work zones and crashes. A key feature of the Gen II sensor package is providing layers of redundancy to one another. The perception algorithms frequently cross-check the data from different sensors to increase the chances that no object is left undetected and to eliminate false positives. This sensor package manifests as mature sensor fusion technology able to combine the sensing capabilities of multiple sensors, resulting in more reliable and robust perception with a broad sensing scope. To this end, the underlying perception algorithms for processing the data are more advanced and capable of performing complex sensor fusion calculations and applying artificial intelligence algorithms in real time. Another envisioned key feature of this generation of sensors is integrating V2V and V2I communication within the vehicles through the OBU. This provides opportunities for the vehicle to receive real-time dynamic data for weather, work zones, and traffic. Risk Assessment Risks • There is a lack of standardization in vehicles and among service providers. • At low market penetration rates, fleet-operated ADS-DVs could emerge and contribute to new crash types. Specifically, at low penetration rates, traditional nonautomated vehicles following fleet-operated ADS-DV could experience increased risk for rear-end crashes. Expected Timeline Vehicle Type Sensor Package Key Infrastructure Requirements Digital Physical Short term High disruption. Light duty. Gen I. GPS-relevant infrastructure. • Pavement markings in excellent condition. • Signage in excellent or good condition. • Parking infrastructure with charging infrastructure. Medium term High disruption. Light duty. Gen II. • GPS-relevant infrastructure. • V2V- and V2X- relevant infrastructure. • Weather data. • Work zone data. • Dynamic traffic data. • Lane and pavement markings in excellent condition. • Signage in excellent or good condition. • Curbside management. • Parking infrastructure with charging infrastructure. Table B-34. Sensor specifications and infrastructure requirements of fleet-operated ADS-DV scenarios.

Additional Example Scenarios B-37   • Another concern with the implementation of fleet-operated ADS-DVs could be the expected degradation of some driving skills—depending on the ADS feature—due to reliance on auto- mation and lack of exposure to DDT. Skill degradation is not limited to psychomotor skills— those pertaining to performing maneuvers—but also include decision-making skills (Miller and Parasuraman, 2007). At higher market penetration of fleet-operated ADS-DVs (Level 4 ADS features), state agencies may consider enacting legislation that necessitates the reassess- ment of driving skills for those relying on fleet-operated ADS-DVs for prolonged time with no exposure to driving. • Interacting with first responders may present challenges. Typically, first responders need to be trained on new procedures for interacting with these vehicles, particularly during emer- gency situations. First responders may need to disable, access, or move low-speed shuttles, direct traffic (e.g., signaling right of way), and identify ADS-related hazards. At early stages of deployment, these procedures are expected to vary between manufacturers, and having a plan for these interactions can help address this risk. • Traffic congestion may occur near curbsides and zones specified for loading, pick-up, and dropping. • Fleet-operated ADS-DVs have the potential to reduce public transport and other mode shares (bicycles and pedestrians). • Major investment in electrification infrastructure is needed. Battery limitations in fleet-operated ADS-DV—they are envisioned to operate as electric vehicles—produce range anxiety and limit operational capacity. • Until the technology is fully mature, and ADS-DVs reach high market penetration values, it is speculated that these vehicles will keep moving slower than an average human driver. This would cause increased traffic congestion, especially during rush hours, and could result in more aggressive and frequent lane-change maneuvers (e.g., nonautomated vehicles passing slow-moving robo-taxis). Opportunities • The high-tech vision system has the potential to outperform human drivers in detecting potential safety issues. Unlike distracted or drunk drivers, fleet-operated ADS-DVs operate at their maximum ability. Therefore, they are likely to reduce crash frequency and severity. This would translate into enhanced pedestrian safety at pedestrian crossings and crosswalks. In fact, there is a concern that some pedestrians could exploit this benefit and “harass,” or at least illegally hinder, fleet-operated ADS-DVs. • Reduced traffic violations may have an aggregate positive impact on traffic safety. There is the potential to redirect law enforcement efforts toward the enforcement of pedestrian and bicyclist laws and related behaviors. • The value of time or travel time may be reduced due to fleet-operated ADS-DVs, because of the increased comfort, reduced stress, and increased productivity while traveling as a passenger instead of as a driver (e.g., Childress et al., 2015). • The need for parking spots (on-street and off-street) may be reduced. • Curbs may be reinvented curbs by changing some on-street parking to pick-up, drop-off, transit stops, or commercial use during the different time of the day. • The infrastructure necessary to price the curb dynamically and collect revenues needs develop- ment (for instance, by installing smart meters that display current prices, accept payments, and notify servers if they are occupied). • Fleet-operated ADS-DVs are the among ADS features that provide massive opportunities for better fuel saving and longer vehicle life. Since these vehicles are envisioned to operate as electric vehicles, major reductions in GHG emissions and fuel consumption are expected to occur simultaneously with their deployment. Also, automating the vehicle in congested

B-38 Framework for Assessing Potential Safety Impacts of Automated Driving Systems conditions improves the vehicle’s energy efficiency and motor lifetime. Funding recipients of the DOE NEXTCAR program are commercializing these types of technologies (ARPA-E, 2020). Additionally, smart eco-routing algorithms could be integrated with these vehicles, such that they are capable of finding faster routes to their destinations, drive more efficiently, and consume less fuel. • With increased market share, fleet-operated ADS-DVs provide potential for implementing innovative eco-driving and speed harmonization techniques, when such vehicles are con- nected (supporting V2V or V2I communication systems). More sophisticated eco-driving and speed harmonization techniques could be considered, such as geofencing a cordon that may be scalable and movable. Goals and Hypothesis. For the fleet-operated ADS-DV deployment scenario, the goal is to reduce the frequency and severity of crashes involving on-demand and demand response mobility services in controlled environments and urban centers through use of fleet-operated ADS-DVs (SAE Level 3 and Level 4) and supporting infrastructure. The overall hypothesis is that fleet-operated ADS-DVs will improve safety in controlled envi- ronments and urban areas by reducing the frequency and severity of crashes involving demand response and on-demand mobility vehicles within 3 to 5 years. The expected change in the number and percentage of crashes involving demand response and on-demand mobility vehicles will depend on market penetration, which is explored in the analysis. The questions related to the hypothesis are listed below and relate to crash types, crash severity levels, infrastructure, and data. 1. How will the frequency of crashes involving on-demand mobility vehicles change in relation to safety? It is anticipated that crashes involving on-demand mobility service vehicles will decrease with the deployment of fleet-operated ADS-DVs by removing human error from the vehicles. 2. How will the severity of crashes involving on-demand mobility vehicles change in relation to safety? 3. How will the frequency of crashes involving on-demand mobility service vehicles (traditional human-operated vehicles) change in relation to safety? There is the potential for road users to change travel modes from driving their own vehicles to using fleet-operated ADS-DVs. Therefore, this switch would impact crashes involving regular user-owned vehicles. 4. Will the safety and deployment of fleet-operated ADS-DVs change if the ODD is expanded and/or infrastructure improvements are made? The current ODD for fleet-operated ADS-DVs includes certain road types, environmental conditions, and infrastructure conditions (e.g., pavement markings in excellent conditions, electric vehicle charging stations, V2X-related infrastructure). If fleet-operated ADS-DVs can operate in additional conditions, then there is potential to expand the safety benefits. 5. Will frequency and severity of pedestrian- and bicycle-related crashes change with the deployment of fleet-operated ADS-DVs? It is anticipated that pedestrian crashes will decrease in frequency and severity. This is due to the deployment of fleet-operated ADS-DVs (i.e., removing human error from the vehicles) in controlled and urban environments where pedestrians are typically present. In summary, through the deployment of fleet-operated ADS-DVs, it is hypothesized that the frequency and severity of crashes involving demand response vehicles, on-demand mobility vehicles, and individual-owned vehicles, including pedestrian- and bicycle-related crashes, will be reduced. While this provides an overview of potential safety impacts, it is important to perform a crash sequencing exercise to think through the contributing factors and precipitating events that lead to a crash. The following are a few examples related to ADS-DVs:

Additional Example Scenarios B-39   1. Rear-end: a. Congested conditions cause vehicle 1 to stop. Driver of vehicle 2 (a fleet-operated vehicle) is distracted, falls asleep, or is otherwise inattentive. Vehicle 2 rear-ends vehicle 1. 2. Right-angle: a. Driver of vehicle 1 (a fleet-operated vehicle) stops at a stop-controlled approach of a two- way stop-controlled intersection. Vehicle 2 is traveling on the uncontrolled approach. Driver of vehicle 1 does not see vehicle 2 because of sun glare. Vehicle 1 proceeds through the intersection. Vehicle 2 collides with vehicle 1. 3. Run-off-road: a. Driver of vehicle 1 (a fleet-operated vehicle) is distracted, falls asleep, or is otherwise inattentive and the vehicle drifts off the road. b. Driver of vehicle 1 (a fleet-operated vehicle) is fully attentive and adverse weather con- tributes to driver losing control or incorrectly navigating and the vehicle leaves the road. From the anticipated capabilities of ADS-DVs and the above crash sequencing, the research team identified specific opportunities for this ADS application to mitigate crashes. For example, ADS-DVs are not expected to operate in adverse conditions, so there is limited potential to mitigate crashes related to sequence 3b; however, ADS-DVs are expected to provide opportunities to mitigate crashes related to sequence 1a, 2a, and 3a among others. Methods of Analysis Components of the Analysis To estimate the safety effects of fleet-operated ADS-DV, an analysis is needed to quantify the potential safety impacts using crash data and the types of crashes fleet-operated ADS-DVs are expected to influence, including the operating environments of the expected deployment. The ODD of fleet-operated ADS-DVs is separated by area type, speed conditions, environmental conditions, and traffic features. For the short- and medium-term timelines, fleet-operated ADS-DVs can operate only in sub- urban geofenced areas. However, fleet-operated ADS-DVs in the medium-term timeline can also operate in urban geofenced areas. While there are some pilot projects of ADS-DV in urban areas, these services are not available as fully automated options and are not expected to be available on a larger scale in the short term. The streets and intersections in the short-term time- line are similar to those in suburban areas, where speeds can reach up to 45 mph. The streets and intersections in the medium-term timeline are similar to those of suburban areas and central business districts, and speeds can reach up to 50 mph. To operate in these area types, the roadways must have well-marked and well-maintained signs and pavement markings in excellent condi- tions for both the short- and medium-term timelines. Fleet-operated ADS-DVs are anticipated to only operate in clear weather conditions for the short-term timeline. This includes no rain, snow, or ice. Fleet-operated ADS-DVs can similarly only operate in clear weather conditions in the medium-term timeline; however, they can also operate in light or mild rain conditions. Fleet-operated ADS-DVs are expected to affect specific crash types and crashes with specific types of vehicles. The crash types typically include those relating to the types of vehicles fleet- operated ADS-DVs will replace. These include crashes with on-demand mobility service vehicles, demand response vehicles, and privately owned vehicles where users switched travel modes to fleet-operated ADS-DVs. The use of fleet-operated ADS-DVs will also potentially affect crashes where the crash contributing factor relates to driver behavior (e.g., speeding) or condition (e.g., distraction, impaired) due to the removal of human error from the driving scenario with the use of fleet-operated ADS-DVs.

B-40 Framework for Assessing Potential Safety Impacts of Automated Driving Systems A summary of the components of the analysis is shown in Table B-35. It is assumed that fleet-operated ADS-DVs will not eliminate all crashes, including pedestrian and bicycle crashes, where the vehicles involved are vehicles that fleet-operated ADS-DVs are expected to replace. The assumption is that they will eliminate only a certain percentage of crashes because there remains the potential for road-user error or technology failure for fleet- operated ADS-DVs. The expected timeline for achieving success depends on the timeline for deployment and penetration rates of fleet-operated ADS-DVs. In this scenario, the expected timeline is assumed to pertain to a high disruption of fleet-operated ADS-DVs. Analysis Methods Data Sources The source of crash data used to estimate the effect of fleet-operated ADS-DVs on safety is the National Transit Database (NTD). The NTD, provided by the U.S. Department of Transportation Federal Transit Administration (FTA), offers crash information related to transit vehicles. Transit agencies in urban areas (850 transit providers) report financial and operating information to the FTA, where it is compiled and offered to the public. Some of the information available includes agency, transit mode, location (city and state), directional route miles, passenger miles traveled, events, collisions, fatalities, and injuries. By law, California mandates all companies that are actively testing Level 4 driverless vehicles on California public roads to disclose the number of miles driven and the frequency in which human safety drivers were forced to take control of their autonomous vehicles (Autonomous Vehicle Disengagement Reports). Manufacturers are also required to provide the California Department of Motor Vehicles with Traffic Collision Involving an Autonomous Vehicle report (form OL 316) within 10 days after the collision (Autonomous Vehicle Collision Reports). These reports could provide valuable data for refining the framework results, once these driverless vehicles are commercially deployed at wide scale. Component Description Functional classification, lanes, and road configuration. All streets and intersections within area type. Area type. Suburban geofenced area. Urban geofenced area. Operating speed. Up to 45 mph (short term). Up to 50 mph (medium term). Weather conditions. Clear (short term). Clear and mild and light rain (medium term). Crash types. On-demand mobility service vehicles. Transit vehicles. Privately owned vehicles. Pedestrian and bicycle. Crash severity level. All severity levels. Table B-35. Components of the analysis.

Additional Example Scenarios B-41   Define Metrics The primary metric for assessing the safety impact of fleet-operated ADS-DVs is the change in crashes by frequency. While the change in crashes should be based on the crashes of interest for the deployment scenario, the NTD data are not sufficiently detailed to categorize and filter crashes based on area type, infrastructure conditions, speed requirements, and weather conditions. Therefore, the data were filtered based on passenger miles traveled and mode. The passenger miles traveled were required to be greater than 0 miles to ensure availability of a normalization parameter for the crashes. As mentioned, the data were also filtered by mode. The modes included from the NTD used in the analysis are demand response (including taxi), jitney, and vanpool. Evaluation Method It is important to test the hypothesis and related questions based on the available data. The evaluation uses crash history and identifies the crashes that could potentially be affected by fleet- operated ADS-DVs. The evaluation estimates crash reduction by frequency. Table B-36 lists the categories of crashes used to estimate the safety impact of fleet-operated ADS-DVs. For the analysis, the collisions are normalized using the passenger miles traveled to obtain a rate (i.e., collision, fatality, or injury rate), shown in the following equation per 100 million passenger miles traveled. =Rate number of events passenger miles traveled miles100,000,000 Using the normalized numbers, the estimated reductions in related events (collisions with vehicle, pedestrian, or bicyclist) were calculated based on the percentage of VMT share repre- sented by the fleet-operated ADS-DVs. For example, rates calculated using the above equation were multiplied by the VMT share, shown in the equation below. = ×Reduction rate VMT share A sensitivity analysis is used to explore various assumptions related to penetration rates and probabilities that a vehicle is a fleet-operated ADS-DV and the technology functions properly. Category Event Type or Severity 1 Collision with vehicle. 2 Collision with person. 3 Total events (including collisions and non-collisions). 4 Fatalities―bicyclist. 5 Fatalities―pedestrian. 6 Total fatalities. 7 Injuries―bicyclist. 8 Injuries―pedestrians. 9 Total injuries. Table B-36. Crash types and severity levels to estimate fleet-operated ADS-DV impact.

B-42 Framework for Assessing Potential Safety Impacts of Automated Driving Systems Results Statistics Table B-37 through Table B-39 display crash statistics for the NTD data for 2008 to 2018. The crashes are separated by events and collisions, fatalities, and injuries, respectively. Impact of Fleet-Operated ADS-DVs The potential crash rate reduction is based on the VMT share described in the deployment scenarios. The VMT shares for the different timelines of fleet-operated ADS-DVs are listed below. Short-term (high disruption) VMT share = 10% Medium-term (high disruption) VMT share = 22% In addition to the percentage of VMT share shown above, other percentages of VMT share were used as a sensitivity analysis to account for fleet-operated ADS-DVs that do not have the auto- nomous feature engaged or for situations when the technology fails. Given the wide range of projec- tions for market penetration, 50% of the VMT share (listed above) were assumed as the values for the sensitivity analysis, listed below. This helps to show the relative difference in results if only half of the expected market penetration is achieved. Collision Type Number of Collisions Collisions per 100 Million Passenger Miles Traveled Collision with vehicle 6,200 26.74 Collision with person 463 2.00 Total collisions and non-collisions 16,456 70.97 Table B-37. Number of collisions and collision rate. Fatality Type Number of Fatalities Fatalities per 100 Million Passenger Miles Traveled Fatalities―bicyclist 0 0.00 Fatalities―pedestrians 9 0.04 Fatalities total 92 0.40 Table B-38. Number of fatalities and fatality rate. Injury Type Number of Injuries Injuries per 100 Million Passenger Miles Traveled Injuries―bicyclist 138 0.60 Injuries―pedestrians 334 1.44 Injuries total 19,272 83.11 Table B-39. Number of injuries and injury rate.

Additional Example Scenarios B-43   Short-term VMT share for sensitivity analysis = 5% Medium-term VMT share for the sensitivity analysis = 11% The estimated reductions for total collisions, fatalities, and injuries were calculated using the method previously described, where the rate is multiplied by the VMT share. The analysis considered pedestrian and bicyclist fatalities and injuries separately as well as collision with person. However, the data for “collision with person” did not distinguish between collision with a bicyclist and collision with a pedestrian. The analysis assumed collision with a person refers to both a bicyclist and pedestrian. Table B-40 displays the reduction in rates for the short-term timeline due to the deployment of fleet-operated ADS-DVs. Table B-41 displays the reduction in rates for the medium-term timeline due to the deployment of fleet-operated ADS-DVs. In both tables, the rate reduction is shown in terms of events per 100 million pas- senger miles traveled. Interpreting the Results The reductions in Table B-40 and Table B-41 indicate that events involving demand response— taxis, jitney, and vanpool vehicles, which are used as surrogates for fleet-operated ADS-DVs— are expected to decrease with the deployment of fleet-operated ADS-DVs. As time progresses from the short-term to medium-term timeline, a greater reduction is expected. Agencies can use the results to estimate the potential reduction in events. The following example displays rate reduction for a given agency. The agency is considering conversion of a fleet of traditional, driver-operated, demand-response vehicles to fleet-operated ADS-DVs. Based on historical data, there are 80 events involving the traditional, driver-operated, demand-response vehicles on routes with an associated 97,249,449 passenger miles traveled. The calculated event rate is 82.26 events per 100 million passenger miles traveled. Using the short- and medium- term VMT share, the reduction in crash rate and new crash rates are calculated, as shown in Table B-42. Assuming a short-term VMT share of 5%, the expected reduction is 4 events for Fatalities―pedestrians 0.002 0.00 Fatalities total 0.02 0.04 Injuries―bicyclist 0.03 0.06 Injuries―pedestrians 0.07 0.14 Injuries total 4.16 8.31 Collision/Fatality/Injury Type Rate Reduced (5% VMT Share) (Events per 100 Million Passenger Miles Traveled) Rate Reduced (10% VMT Share) (Events per 100 Million Passenger Miles Traveled) Collision with vehicle 1.34 2.67 Collision with person 0.10 0.20 Total collisions and non-collisions 3.55 7.10 Fatalities―bicyclist 0.00 0.00 Table B-40. Crash rate reduction for the short-term timeline.

B-44 Framework for Assessing Potential Safety Impacts of Automated Driving Systems every 100 million passenger miles traveled. An agency could perform a simple or more detailed benefit-cost analysis to compare these benefits with the cost to convert the vehicle fleet as part of the decision process. Communicate Outcomes Through the framework and analysis, the results demonstrate how rates are expected to decrease. Specifically, fleet-operated ADS-DVs are expected to reduce crashes. If the ODD is extended so fleet-operated ADS-DVs can operate in different areas, different facilities, or different environmental conditions, a further reduction in crashes would be expected. However, this could not be verified in the analysis because of the limited variables available in the dataset. Assumptions were made to perform the analysis. As previously mentioned, the exact elements from the ODD (e.g., area type and weather) could not be filtered or separated. For example, fleet- operated ADS-DVs can only operate in specific area types. The crashes could not be separated by only those occurring in specific locations. Therefore, the analysis assumed the crashes were occurring in locations where fleet-operated ADS-DVs would operate. Reduction in Rate (Events per 100 Million Passenger Miles Traveled) New Rate (Events per 100 Million Passenger Miles Traveled) Short term: • VMT share = 5% 4.11 78.15 • VMT share = 10% 8.23 74.04 Medium term: • VMT share = 11% 9.05 73.21 • VMT share = 22% 18.10 64.16 Table B-42. Example of crash reduction. Collision/Fatality/Injury Type Rate Reduced (11% VMT Share) (Events per 100 Million Passenger Miles Traveled) Rate Reduced (22% VMT Share) (Events per 100 Million Passenger Miles Traveled) Collision with vehicle 2.94 5.88 Collision with person 0.22 0.44 Total collisions and non-collisions 7.81 15.61 Fatalities―bicyclist 0.00 0.00 Fatalities―pedestrians 0.004 0.01 Fatalities total 0.04 0.09 Injuries―bicyclist 0.07 0.13 Injuries―pedestrians 0.16 0.32 Injuries total 9.14 18.29 Table B-41. Crash rate reduction for the medium-term timeline.

Additional Example Scenarios B-45   Another assumption used in the analysis relates to the environmental conditions. Fleet-operated ADS-DVs can only operate in clear and windy weather conditions for the short-term timeline and in weather conditions consisting of no heavy rain or snow for the medium-term timeline. However, the weather conditions for the crashes in the NTD are unknown. The analysis assumed that the crashes occurred in weather conditions in which fleet-operated ADS-DVs can operate in the short- and medium-term timelines. Other elements of fleet-operated ADS-DVs that were assumed to be present in the data for the short- and medium-term timelines include well-marked and well-maintained roads, parking infrastructure with electric vehicle charging stations, and GPS-relevant infrastructure.