Autonomous Vehicles

Autonomous vehicle (AV) technology will fundamentally change transportation. Technological advancements are creating a continuum between conventional, fully human-driven vehicles and AVs which partially or fully drive themselves and may ultimately require no driver at all. Within this continuum are technologies that enable a vehicle to assist and make decisions for a human driver. Such technologies include crash warning systems, adaptive cruise control (ACC), lane keeping systems, and self-parking technology.

Equipping cars and light vehicles with this technology will reduce crashes, energy consumption, and pollution—and reduce the costs of congestion.

This technology is most easily conceptualized using a five-level hierarchy suggested by the National Highway Traffic Safety Administration (NHTSA) with different benefits realized at different levels of automation:

  • • Level 0 (no automation): The driver is in complete and sole control of the primary vehicle functions (brake, steering, throttle, and motive power) at all times and is solely responsible for monitoring the roadway and for safe vehicle operation.
  • • Level 1 (function-specific automation): Automation at this level involves one or more specific control functions. If multiple functions are automated, they can operate independently of each other. In this case, the driver has overall control and is solely responsible for safe operation but can choose to cede limited authority over a primary control (as in ACC). Alternatively, the vehicle can automatically assume limited authority over a primary control (as in electronic stability control), or the automated system can provide added control to aid the driver in certain normal driving or crash-imminent situations (e.g., dynamic brake support in emergencies).
  • • Level 2 (combined function automation): This level involves automation of at least two primary control functions designed to work in unison to relieve the driver of controlling those functions. Vehicles at this level of automation can share authority when the driver cedes active primary control in certain limited driving situations. The driver is still responsible for monitoring the roadway and safe operation and is expected to be available for control at all times and on short notice. The system can relinquish control with no advance warning, and the driver must be ready to control the vehicle safely.
  • • Level 3 (limited self-driving automation): At this level of automation, the driver can cede full control of all safety-critical functions under certain traffic or environmental conditions and rely heavily on the vehicle to monitor changes in those conditions requiring transition back to driver control. The driver is expected to be available for occasional control but with sufficiently comfortable transition time.
  • • Level 4 (full self-driving automation): The vehicle is designed to perform all safety-critical driving functions and monitor roadway conditions for an entire trip. Such a design anticipates that the driver will provide destination or navigation input but is not expected to be available for control at any time during the trip. This includes both occupied and unoccupied vehicles.

The type and magnitude of the potential benefits of AV technology will depend on the level of automation achieved. For example, some of the safety benefits of AV technology may be achieved from function-specific automation (e.g., automatic braking), while the land use and environmental benefits are likely to be realized only by full automation (Level 4) (Anderson et ah, 2016).

1.1.1 Brief History and Current State of Autonomous Vehicles

For decades, futurists have envisioned vehicles that drive themselves, and research into AV technology can be divided into three phases. Phase 1: Foundational Research

From approximately 1980 to 2003, university research centers worked on two visions of vehicle automation. As one thrust, researchers pursued the development of automated highway systems (AHS), in which vehicles depended significantly on the highway infrastructure to guide them (Anderson et ah, 2016).

One of the first major demonstrations of such a system took place in 1997 over a 7.6-mile stretch of California’s 1-15 highway near San Diego. Led by the California Partners for Advanced Transit and Highways (PATH) program, the “DEMO 97” program demonstrated the platooning of eight AVs guided by magnets embedded in the highway and coordinated with vehicle-to-vehicle (V2V) communication (Ioannou, 1998).

A second research thrust was to develop both semiautonomous and AVs that depended little, if at all, on highway infrastructure. In the early 1980s, a team led by Ernst Dickmanns at Bundeswehr University Munich in Germany developed a vision-guided vehicle that navigated at speeds of 100 km/h without traffic (Baela & Maarton, 2011). Carnegie Mellon University’s NavLab developed a series of vehicles, named NavLab 1 through NavLab 11, from the mid-1980s to the early 2000s. In July 1995, NavLab 5 drove across the country in a “No Hands Across America” tour, in which the vehicle steered autonomously 98% of the time, while human operators controlled the throttle and brakes. Other similar efforts around the world sought to develop and advance initial AV and highway concepts (Anderson et al., 2016). Phase 2: Grand Challenges

From 2003 to 2007, the US Defense Advanced Research Projects Agency (DARPA) held three “Grand Challenges” that markedly accelerated advancements in AV technology. The first two Grand Challenges charged research teams with developing fully autonomous vehicles for competition in a 150-mile off-road race for $1 million and $2 million prizes, respectively. No vehicle completed the 2004 Grand Challenge— the best competitor completed less than 8 miles of the course (BBC News, 2004). However, five teams completed the 2005 Grand Challenge course, held only 18 months later. The fastest team completed the course in just under 7 h, with the next three fastest finishing within the next 35 min (DARPA, undated).

In 2007. DARPA held its third and final AV challenge, dubbed the “Urban Challenge.” As the name suggests, vehicles raced through a 60-mile urban course, obeying traffic laws and navigating alongside other autonomous and human-driven vehicles. Six teams finished the course, and three completed the race within a time of 4.5h, including time penalties for violating traffic and safety rules. This Grand Challenge spearheaded advancements in sensor systems and computing algorithms to detect and react to the behavior of other vehicles, to navigate marked roads, and to obey traffic rules and signals (Anderson et al., 2016). Phase 3: Commercial Development

The DARPA Challenges solidified partnerships between auto manufacturers and the education sector and mobilized a number of endeavors in the automotive sector to advance AVs. These included the Autonomous Driving Collaborative Research Lab, a partnership between GM and Carnegie Mellon University (Carnegie Mellon University, undated) and a partnership between Volkswagen and Stanford University (Stanford University, undated).

Google’s Driverless Car initiative has brought autonomous cars from the university laboratory into commercial research. The program began shortly after the DARPA Urban Challenge and drew on the talents of engineers and researchers from several teams who participated in that competition. In the years since, Google has developed and tested a fleet of cars and initiated campaigns to demonstrate the applications of the technology through, for example, videos highlighting mobility offered to the blind (Google, 2012). Google is not alone. In 2013, Audi and Toyota both unveiled their AV visions and research programs at the International Consumer Electronics Show, an annual event held every January in Las Vegas (Hsu, 2013). Nissan has also recently announced plans to sell an AV by 2020 (Anderson et al„ 2016).

1.1.2 Current State of Autonomous Vehicle Technology

Google’s vehicles, operating fully autonomously, have driven more than 500,000 miles without a crash attributable to the automation. Advanced sensors to gather information about the world, increasingly sophisticated algorithms to process sensor data and control the vehicle, and computational power to run them in real time have permitted this level of development (Anderson et al., 2016). Making Sense of the World

In the most general terms, AVs employ a “sense-plan-act” design that is the foundation of many robotic systems. A suite of sensors on the vehicle gathers raw data about the outside world and the vehicle’s relations to its environment. Software algorithms interpret the sensor data—for example, lane markings from images of the road or behavior of other vehicles from radar data. They use these data to make plans about the vehicle’s own actions—its overall trajectory down the road and immediate decisions such as accelerating and changing directions. These plans are converted into actionable commands to the vehicle’s control system; i.e., steering, throttle, brakes.

Many “sense-plan-act” loops run in parallel on an AV. One loop may run at extremely high frequency to initiate rapid emergency braking, while another runs less frequently to plan and execute complex behaviors such as changing lanes. In some cases, the planning component of the loop is extremely short and resembles a sense-act cycle instead of a sense-plan-act cycle. For instance, a vehicle may gather data about obstacles immediately in front of it at very high frequency and initiate emergency braking if any obstacle is detected within a short distance. In this case, the sensor data may directly trigger a vehicle action.

With perfect perception (a combination of sensor data gathering and interpretation of those data), AVs could plan and act perfectly, achieving ultra-reliability. Vehicles never tire; their planning algorithms can choose provably optimal behaviors, and their execution can be fast and flawless. For example, if a deer were to leap into the path of a human-driven vehicle, the driver may make mistakes in choosing whether to swerve, brake, or take another course of action. The driver may also make mistakes in executing the action; for example, oversteering a swerve. AVs need never make these mistakes. Computer algorithms can rapidly evaluate, compare, select, and execute the best action from among a number of maneuvers, taking into account the vehicle’s speed, the animal’s trajectory, the position and behavior of other vehicles, and the utility of various outcomes (Anderson et al., 2016).

One of the more difficult challenges for AVs is making sense of the complex and dynamic driving environment—for example, perceiving the deer. The driving environment includes many elements (Anderson et al., 2016):

  • • Other vehicles on the road, each of which operates dynamically and independently.
  • • Other road users or on-road obstacles, such as pedestrians, cyclists, wildlife, and debris.
  • • Weather conditions, from sunny days to severe storms.
  • • Infrastructure conditions, including construction, rough road surfaces, poorly marked roads, and detours.
  • • Traffic events, such as congestion or crashes.

It is in making sense of the world that humans often outperform robots. Human eyes are sophisticated and provide nearly all of the sensory data we use to drive. We are also adept at interpreting what we see. Although our eyes are passive sensors, only receiving information from reflected light, we can judge distances, recognize shapes, and classify objects such as cars, cycles, and pedestrians and see in a tremendous range of conditions. Of course, we are far from perfect. Our sight and our cognition of visual information vary and can be dangerously limited in several situations: adverse ambient conditions such as darkness, rain, and fog, when we are tired or distracted, or when we are impaired through the use of drugs or alcohol (Olson, Dewar, & Farber, 2010).

A second limitation is that, like human eyes, camera systems are better able to gather data in some ambient conditions (e.g., clear sunny days) than others (e.g., fog or rainstorms). Changes in ambient conditions also pose challenges, as camera systems calibrated to certain conditions may have difficulty interpreting data in others. This problem of autonomous camera calibration is also a fundamental robotics research problem (Furukawa & Ponce, 2009).

Camera-based systems, i.e., computer vision systems, are analogous to human eyes and visual cognition. They can “see” very long distances and provide rich information about everything in their field of view. Cameras are also inexpensive, making them important components for cost-effective autonomy. However, they have two important limitations. First, the underlying algorithms are not nearly as sophisticated as humans at interpreting visual data. The Solutions in Perception Challenge is an annual competition that embodies this difference, challenging engineering teams to develop computer vision and other sensor algorithms that can detect, recognize, and locate objects. In the 2011 competition, for example, the objects included a number of items that would be found on supermarket shelves. None of the competing teams reached the goal of 80% accuracy (Markoff, 2011).

AVs have a critical advantage over humans: they can draw upon a much wider array of sensor technologies than cameras alone. While many major advances have been made in the last decade, however, the interpretation of visual data (and sensor data more generally) remains a fundamental research problem in the field of computer vision. We can expect advances in both sensor technology and perception algorithms, but matching human perception under best conditions is a long-term research challenge (Anderson et al., 2016). Light Detection and Ranging Sensor Systems

Light detection and ranging (LiDAR) systems feature prominently in robotic systems, including AVs. LiDAR systems determine distances to obstacles by using laser range finders, which emit light beams and calculate the time of flight until a reflection is returned by objects in the environment. Many sophisticated LiDAR systems couple multiple laser range finders with rapidly rotating mirrors to generate 3D point clouds of the environment. Developed during the DARPA Grand Challenges and used by teams in the Urban Challenge and by Google, the Velodyne HDL-64E LIDAR uses 64 lasers that provide 1.3 million data points per second and offer a 360° field of view. LIDAR is typically useful over a shorter range than other sensors—the Velodyne provides data up to 120 m away, depending on the reflectivity of the object.

LiDAR systems’ two key limitations are range (less useful at long ranges) and reflectivity (poor reflection from certain kinds of materials). The Velodyne’s specifications state that it detects black asphalt, which has low reflectivity, to a range of just 50 m (Velodyne, 2010). The costs of LiDAR systems range widely but are expected to decline in the near future (Anderson et al., 2016). Sensor Suites

Each sensor provides different kinds of data and has its own limitations related to field of view, ambient operating conditions, and the elements in the environment that it can sense. Because the limitations are fairly well understood, the usual practice is to construct suites of complementary sensors that are positioned around the vehicle to prevent blind spots—both visual blind spots (i.e., due to occluded views) and material blind spots (i.e., the inability to detect certain kinds of objects or certain properties of objects in the environment).

Sensors can be integrated to perceive more about the environment than can be learned purely from the sum of individual sensors’ data. As one example, vision systems can detect colors of surfaces in the distance, while LiDAR can be used to determine the material as that surface approaches. When these two types of sensors are coupled, a system can learn that green surfaces in the distance correspond to grass, allowing the vehicle to make greater sense of the far away environment (Thrun et al., 2007).

Vehicles also use sensor suites for localization, i.e., determining their own position in the world. The use of the Global Positioning System (GPS) is essential for localization. Vehicle GPSs receive signals from orbiting satellites to triangulate their global coordinates. These coordinates are cross-referenced with maps of the road network to enable vehicles to identify their position on roads.

The accuracy of GPS has improved significantly since 2000, when the US government made it fully available to civilian users. However, GPS error can still be large—several meters, even under ideal conditions. The errors grow rapidly when obstacles or terrain occlude the sky, preventing GPS receivers from obtaining signals from a sufficient number of satellites. This is a significant concern in urban areas, where skyscrapers create “urban canyons” in which GPS availability is severely limited (Anderson et al., 2016).

GPS is typically coupled with inertial navigation systems (INSs), which consist of gyroscopes and accelerometers, to continuously calculate the position, orientation, and velocity of a vehicle without the need for external references. INSs are used to improve the accuracy of GPS and to fill in “gaps” such as those caused by urban canyons. The key challenge with INS is drift; even over very short periods, small errors can aggregate into large differences between calculated and true positions. For example, a 10 s period during which the system relies on INS because the GPS signal is unavailable can result in more than a meter of drift in calculated position, even with some of the most sophisticated systems (Applanix, 2012). Environmental Challenges

Certain ambient conditions (e.g., severe precipitation, dense fog) may pose problems for multiple sensors simultaneously. Common failure conditions such as these limit the extent to which sensor combinations can compensate for individual sensor limitations. It must be noted, however, that these same conditions pose problems for humans. Indeed, robotic sensors such as radar may prove more effective than human vision, and the rapid reaction of planning algorithms may be particularly valuable, making autonomous systems imperfect but potentially safer than human drivers in these adverse conditions (Anderson et al., 2016).

Terrain poses challenges as well. A sensor configuration appropriate for a flat environment may be inappropriate for steep hills, where sensors must look “up” or “down” the slopes. Different terrains can require different sensor configurations, which may not be readily changeable. While sensors can be put on adjustable mounts to accommodate this problem, this adds complexity and cost (Urmson et al., 2006).

Road materials also change from region to region. They are typically concrete and asphalt but can be dirt, cobblestone, and other materials. Different materials have different reflectivity, and sensors calibrated to certain materials may have difficulty detecting other materials with equal fidelity.

Construction projects and roadwork are particularly difficult to negotiate, as there may be little consistency in signage and alerts, roadway materials may change suddenly, and the maneuvers needed to navigate through construction zones may be complex and poorly marked. Moreover, these areas often involve deviations from preconstructed maps, so vehicle localization may be particularly difficult.

Each of these factors can have implications for where AVs can successfully operate. For example, in the United States, weather and terrain vary significantly, as do the road materials and signage practices. A vehicle that operates easily on flat terrain in Louisiana may have significant performance challenges on Colorado’s snowy and steep roads or in New York City’s congested urban canyons (Anderson et al., 2016). Graceful Degradation

Sensor failure (as opposed to external environmental conditions) can pose serious performance threats (Hwang, Kim, Kim, & Eng, 2010). Sensors may fail because of electrical failures, physical damage, or age. It will be critical for AVs to have internal sensing and algorithms that can detect when internal components are not performing adequately. This is not easy. A sensor that fails to provide any data is easily detected as nonfunctioning, but a sensor that occasionally sends spurious data may be much harder to detect.

These and other failures will require a system that degrades gracefully (Berger & Rumpe, 2012). AVs will likely need to have an ultrareliable and simple low-level system that uses minimal sensor data to perform basic functions in the event of main system degradation or failure. The backup system must also be able to detect degradation and failure and override control rapidly and safely. The task of graceful degradation may be complicated by traffic conditions and roadways. If a system fails in the middle of a curve in dense traffic, it may need to be able to navigate to a safe area to pull over (Anderson et al., 2016). V2V and V21 Communication

The role of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication in enabling AV operation remains unclear. While this technology could ease the task of automated driving in many circumstances, it is not clear that it is necessary. Moreover, V2I might require substantial infrastructure investments—for example, if every traffic signal must be equipped with a radio for communicating with cars. Sharing the Drive

Partly as a result of all of these challenges, most stakeholders anticipate a “shared driving” concept will be used on the first commercially available AVs. Vehicles will drive autonomously in certain operating conditions, for example, below a particular speed, only on certain kinds of roads, in certain driving conditions, and will revert to traditional, manual driving outside those boundaries or at the request of a human driver.

To experience the greatest benefits of the technology, human drivers will need to be able to engage in other tasks while the vehicle is driving autonomously. For safety, however, they will need to quickly reengage (in a matter of seconds or less) at the vehicle’s request. Cognitive science research on distracted driving suggests this may be a significant safety challenge. Similarly, developing the appropriate mental models for human-machine collaboration may be a challenge in creating a technology usable by the general public. Integrity, Security, and Verification

Software upgrades might need to be backward-compatible with earlier models of vehicles and sensor systems. Moreover, as more vehicle models offer autonomous driving features, software and other system upgrades will have to perform on increasingly diverse platforms, making reliability and quality assurance even more challenging. System security is also a concern; viruses or malware must be prevented from subverting proper functioning of vehicles’ systems.

State transportation departments may need to anticipate the use of vastly different kinds of AVs operating on roadways. This may pose challenges for the registration and requirements necessary for the vehicles to operate and for the level of training particular operators must have. One short-term action that might improve safety is requiring stricter conformance to road signage requirements, particularly those involving construction or some alteration to the roadway. This would aid human drivers and ease some of the perception requirements for AVs (Anderson et al., 2016).

1.1.3 Why Is Autonomous Vehicle Technology Important Now?

AV technology merits the immediate attention of policymakers for several reasons. First, the technology appears close to maturity and commercial introduction. Google’s efforts—which involve a fleet of cars that collectively have logged hundreds of thousands of autonomous miles—have received widespread media attention and demonstrate this technology has advanced considerably. Every major commercial automaker is engaged in research in this area, and the full-scale commercial introduction of truly autonomous (including driverless) vehicles is predicted to occur within 5-20years. Several states in the United States have passed laws to regulate the use of AVs, and many more laws have been proposed. As these technologies trickle (or flood) into the marketplace, it is important for policymakers to understand the effects that existing policy (or lack thereof) is likely to have on the development and adoption of this technology (Anderson et al., 2016).

Second, the stakes are high. In the United States alone, more than 30,000 people are killed each year in crashes, approximately 2.5 million are injured, and the vast majority of these crashes are the result of human error (Choi, Zhang, Young, Singh, & Chen, 2008). By reducing the opportunity for human error, AV technologies have the potential to reduce the number of crashes (Anderson et al., 2016).

AVs may reduce congestion and its associated costs; estimates suggest the effective road capacity (vehicles per lane per hour) can be doubled or tripled. The costs of congestion can be greatly reduced if vehicle operators can productively conduct other work. AV technology promises to reduce energy use as well; automobiles have become increasingly heavy over the past 20years partly to meet more rigorous crash test standards. If crashes become exceedingly rare events, it may be possible to dramatically lighten automobiles.

In the long run, AVs may improve land use. Quite apart from the environmental toll of fuel generation and consumption, the existing automobile shapes much of our built environment. Its centrality to our lives accounts for the acres of parking in even our most densely occupied cities. With the ability to drive and park themselves at some distance from their users, AVs may obviate the need for nearby parking for commercial, residential, or work establishments, and this, in turn, may enable a reshaping of the urban environment and permit new in-fill development as adjacent parking lots become unnecessary.

Along with these benefits, however, AVs could have many negative effects. By reducing the time cost of driving, AVs may encourage greater travel and increase the total vehicle miles traveled (VMT), leading to more congestion. Urban sprawl may increase if commuters move ever farther away from workplaces. Similarly, AVs may eventually shift users’ preferences towards larger vehicles to permit other activities. In theory, this could even include beds, showers, kitchens, or offices. If AV software becomes standardized, a single flaw might lead to many accidents. Internet-connected systems might be hacked by the malicious. And perhaps the biggest risks are simply unknowable (Anderson et al., 2016).

From seatbelts, to air bags, to antilock brakes, automakers have often been reluctant to incorporate expensive new technology, even if it can save many lives (Mashaw & Harfst, 1990). Navigating the AV landscape makes implementation of these earlier safety improvements appear simple by comparison. Negotiating the risks to reach the opportunities will require careful policymaking (Anderson et al., 2016).

1.1.4 Components of Autonomous Vehicles

The AV system can be divided into four main components, as shown in Figure 1.1. The vehicle senses the world using many different sensors mounted on it. These hardware components gather data about the environment. The information from the sensors is processed in a perception block and turned into meaningful information. A planning subsystem uses the output from the perception block to plan behavior and create short- and long-range plans. A control module ensures the vehicle follows the path provided by the planning subsystem and sends control commands to the vehicle (Kocic, Jovicic, & Drndarevic, 2018).

The first fully autonomous vehicles were developed in 1984 at Carnegie Mellon University and in 1987 by Mercedes-Benz and the University of Munich. Since then, many companies and research organizations have developed prototypes of AVs and are intensively working on the development of full vehicle autonomy (Kocic, Jovicic, & Drndarevic, 2018).

As mentioned previously, a significant event in AV development was the Defence Advanced Research Project Agency’s (DARPA) Grand Challenge events in 2004 and 2005 (Thrun, 2006: Montemerlo, 2006) and Urban Challenge event in 2007 (Buehler, Iagnemma, & Singh, 2009). These events demonstrated machines could independently perform the complex human task of driving. In the 2007 DARPA Urban Challenge, six of 11 AVs in the finals successfully navigated an urban environment to reach the finish line, a landmark achievement in robotics (Kocic, Jovicic, & Drndarevic, 2018).

Current challenges in AVs’ development are scene perception, localization, mapping, vehicle control, trajectory optimization, and higher-level planning decisions. New trends in autonomous driving include end-to-end learning and reinforcement learning (Kocic, Jovicic, & Drndarevic, 2018).

Block diagram of the AV system (Kocic, Jovicic, & Drndarevic, 2018)

FIGURE 1.1 Block diagram of the AV system (Kocic, Jovicic, & Drndarevic, 2018).

1.1.5 Autonomous Vehicle Applications

There has been a continuous and gradually increasing interest in developing autonomous ground vehicles for different applications. We classify the types of applications as:

  • 1. Automated highway systems (AHS).
  • 2. In-city and urban driving.
  • 3. Off-road driving.
  • 4. Specialty applications.

The fourth class includes a series of diverse applications and problems, for example, closed deployment environments, tasks requiring special motion, docking, and convoying (Ozguner & Redmill, 2008).

An example of an AHS is shown in Figure l .2. An example of an off-road vehicle and ION, the intelligent off-road navigator, are shown in Figures 1.3a and b, respectively (Chen & Ozguner, 2006; Redmill, Martin, & Ozguner, 2006). Figure 1.3c shows a vehicle designed for city urban driving and developed for DARPA Urban Challenge 2007 (Ozguner & Redmill. 2008).

Studies on the development of AHS usually advocate an ingress-egress pairing, during which the car will follow the assigned lane on the highway. Early testing and demonstration implementations assumed cars would follow specialized technological aids indicating the precise location of the car with respect to the lane.

Figure 1.2 shows a technology advocated for location information with respect to the lane, a radar- reflecting stripe indicating the distance from the center of roadway, and the relative orientation of the car. When this technology was developed, GPS and precision maps were not commonly available. Today, it is assumed that precision maps will identify individual lanes and GPS reception will provide precise location information in real time (Ozguner & Redmill, 2008).

The cars shown in Figure 1.2 are from Demo’97, a test held on a 7.5-mile segment of Highway 1-15 in San Diego. This segment was a segregated two-lane highway normally used for rush-hour high-occupancy vehicle traffic. Traffic flowed in the same direction in both lanes, and there were no intermediate entry and exit points. The curvature of the highway lanes was benign and suited for high-speed (70 mph) driving; other traffic was minimal to nonexistent. A general AHS would presumably have merge and exit lanes, but the single entry-exit aspect of Demo’97 made it a single activity: drive down the lane and possibly handle

Two AVs in Demo’97 following a radar-reflecting stripe and undertaking a pass (Ozguner & Redmill, 2008)

FIGURE 1.2 Two AVs in Demo’97 following a radar-reflecting stripe and undertaking a pass (Ozguner & Redmill, 2008).

(a) Off-road TerraMax at Grand Challenge 2004, (b) Off-road ION at Grand Challenge 2005, (c) Urban vehicle ACT at Urban Challenge 2007 (Ozguner & Redmill, 2008)

FIGURE 1.3 (a) Off-road TerraMax at Grand Challenge 2004, (b) Off-road ION at Grand Challenge 2005, (c) Urban vehicle ACT at Urban Challenge 2007 (Ozguner & Redmill, 2008).

simple interactions with other vehicles. We call this behavior a meta-state. Dealing with interchanges produced by entry and exit lanes would require other meta-states (Ozguner & Redmill, 2008).

The DARPA Grand Challenges of 2004 and 2005 (mentioned above and shown in Figure 1.3a and b) were both off-road races. As such, the only behavior and, thus, the only meta-state required was following the path with obstacle avoidance from point A to point B. However, since there was no “path” or “lane” that could be discerned from a roadway, the only method of navigation was to rely on GPS and INS-based vehicle localization and a series of predefined “waypoints”. Obstacle avoidance was needed, as in an AHS, although in the less structured off-road scenario, greater freedom of movement and deviations from the defined path are allowed. The Grand Challenge race rules ensured there were no moving obstacles, and different vehicles would not encounter each other in motion. General off-road driving would not have this constraint (Ozguner & Redmill, 2008).

Finally, fully autonomous urban driving introduces a significant number of meta-states, situations where different behavior is required, and different classes of decisions need to be made. The DARPA Urban Challenge (mentioned above), although quite complex, had fairly low speed limits, careful drivers, and no traffic lights. Visual lane markings were unreliable, and, thus, true to life, and the terrain was fairly flat, although some areas were unpaved, generating an unusual amount of dust and creating problems for some sensors.

Although “lanes” are obvious in highway systems and urban routes, it is reasonable to assume that off-road environments also present a set of constraints that indicate the drivability of different areas and, thus, provide the possibility of defining lanes. A sketch of a vehicle on a path with waypoints and lanes is shown in Figure 1.4 (Ozguner & Redmill, 2008).

Roadway with waypoints and lane markings (Ozguner & Redmill K, 2008)

FIGURE 1.4 Roadway with waypoints and lane markings (Ozguner & Redmill K, 2008).

1.1.6 Architecture and Hierarchy Issues of Autonomous Vehicles

A generic functional architecture for an AV is given in Figure 1.5 (Ozguner & Redmill, 2008).

The “plan” can be simple or complex, but the overall configuration will cover all three application areas under consideration. Figures 1.6 and 1.7 show the details of the hardware architecture for AVs developed in 1996 and 2007, more than lOyears apart. Although some technologies have changed, and in spite of one being for AHS and the other for autonomous urban driving, the similarities between the two configurations are obvious (Ozguner & Redmill, 2008).

Note that the Demo’97 car does not have a GPS system and relies totally on infrastructure-based queues to find its position with respect to the roadway. The car has a special (stereo) radar system that senses the radar-reflective stripe it straddles on the lane and a vision system that senses the white lane markers on both sides of the lane. It senses other cars on the roadway via radar and a separate LiDAR unit.

Generic functional architecture for an automated vehicle (Ozguner & Redmill, 2008)

FIGURE 1.5 Generic functional architecture for an automated vehicle (Ozguner & Redmill, 2008).

Architecture for an AV used in Demo’97 (Ozguner & Redmill, 2008)

FIGURE 1.6 Architecture for an AV used in Demo’97 (Ozguner & Redmill, 2008).

In the Urban Challenge car developed by OSU, shown in Figure 1.7, the overall architecture is very similar. In this case, direct sensor-based lane detection is not fully integrated into the vehicle, although the sensor suite utilized would have allowed this. The vehicle relied on high-precision GPS signals and inertial and dead-reckoning positioning technologies (Ozguner & Redmill, 2008).

1.1.7 Potential Impacts of Autonomous Vehicles

AV operations are inherently different from human-driven vehicles. AVs can be programmed to not break traffic laws. They do not drink and drive. Their reaction times are quicker, and they can be optimized to smooth traffic flows, improve fuel economy, and reduce emissions. They can deliver freight and unlicensed travelers to their destinations. This section examines some of the largest potential benefits (Fagnant & Kockelman, 2015). Safety

AVs have the potential to dramatically reduce crashes. Table 1.1 highlights the magnitude of automobile crashes in the United States and indicates sources of driver error that may disappear as vehicles become increasingly automated (Fagnant & Kockelman, 2015).

Architecture of ACT. an urban driving car used in DARPA’s Urban Challenge (Ozguner & Redmill. 2008)

FIGURE 1.7 Architecture of ACT. an urban driving car used in DARPA’s Urban Challenge (Ozguner & Redmill. 2008).


US Crash Scope and Selected Human and Environmental Factor Involvement (Fagnant & Kockelman, 2015).

Total crashes per year in the United States

5.5 million

• % human cause as primary factor


Economic costs of US crashes

$277 billion

• % of US GDP


Total fatal and injurious crashes per year in the United States

2.22 million

Fatal crashes per year in the United States

32,367 million

• % of fatal crashes involving alcohol.


• % involving speeding.


• % involving distracted driver.


• % involving failure to keep in proper lane.


• % involving failure to yield right of way.


• % involving wet road surface.


• % involving erratic vehicle operation.


• % involving inexperience or overcorrecting.


• % involving drugs.


• % involving ice, snow, debris, or other slippery surface.


• % involving fatigued or sleeping driver.


• % involving other prohibited driver errors (e.g., improper following, driving on shoulder, wrong side of road, improper turn, improper passing)


Over 40% of these fatal crashes involve some combination of alcohol, distraction, drug involvement, and/or fatigue. Self-driven vehicles would not fall prey to human failings, suggesting the potential for at least a 40% fatal crash-rate reduction, assuming automated malfunctions are minimal and everything else remains constant (such as the levels of long-distance, night-time, and poor-weather driving). Such reductions do not reflect crashes due to speeding, aggressive driving, overcompensation, inexperience, slow reaction times, inattention, and various other driver shortcomings. Driver error is believed to be the main reason for over 90% of all crashes (National Highway Traffic Safety Administration, 2008). Even when the critical reason for a crash is attributed to the vehicle, roadway, or environment, additional human factors such as inattention, distraction, or speeding are regularly found to have contributed to the crash occurrence and/or injury severity (Fagnant & Kockelman, 2015).

The scope of potential benefits is substantial, both economically and politically. Over 30,000 persons die each year in the United States in automobile collisions (National Highway Traffic Safety Administration, 2012), with 2.2 million crashes resulting in injury (Traffic Safety Facts, 2013). Traffic crashes remain the primary reason for the death of Americans between 15 and 24years of age (CDC, 2011).

At $277 billion, the annual economic cost of crashes is over double that of congestion and is highlighted as the number one transportation goal in the nation’s federal legislation. Moving Ahead for Progress in the 21st century (MAP-21). These issues have long been top priorities of the US Department of Transportation’s Strategic Plan.

While many driving situations are relatively easy for an AV to handle, designing a system that can perform safely in nearly every situation is challenging (Campbell, Egerstedt, How, & Murray, 2010). For example, recognition of humans and other objects in the roadway is critical but more difficult for AVs than human drivers (Dalai & Triggs, 2005; Economist Technology Quarterly, 2012; Farhadi, Endres, Hoiem, & Forsyth 2009). A person in a roadway may be small or large, standing, walking, sitting, lying down, riding a bike, and/or partly obscured - all of which complicate AV sensor recognition. Poor weather, such as fog and snow, and reflective road surfaces from rain and ice create other challenges for sensors and driving operations. Evasive decisions should depend on whether an object in the vehicle’s path is a large cardboard box or a large concrete block, and computer vision has much greater difficulty than human vision in identifying material composition. When a crash is unavoidable, it is crucial that AVs recognize the objects in their path so they may act accordingly. Liability for these incidents is a major concern and could be a substantial impediment to implementation (Fagnant & Kockelman, 2015).

Ultimately, some analysts predict that AVs will overcome many of the obstacles that inhibit them from accurately responding in complex environments. Hayes (2011) suggests motor-vehicle fatality rates (per person-mile traveled) could eventually approach those seen in aviation and rail, about 1% of current rates. Some foresee the creation of “crash-less cars” (KPMG & CAR, 2012). However, drivers could take their vehicles out of self-driving mode and take control. Google’s only reported AV crash occurred when a human driver was operating the vehicle. The rate at which human control is needed will be a substantial factor in the safety of these vehicles (Fagnant & Kockelman, 2015). Congestion and Traffic Operations

Aside from making automobiles safer, researchers are developing ways for AV technology to reduce congestion and fuel consumption. For example, AVs can sense and possibly anticipate lead vehicles’ braking and acceleration decisions. Such technology allows smoother braking and fine speed adjustments in following vehicles, leading to fuel savings, less brake wear, and reductions in traffic-destabilizing shockwave propagation. AVs are also expected to use existing lanes and intersections more efficiently with shorter gaps between vehicles, more coordinated platoons, and more efficient route choices. Many of these features, such as ACC, are already being integrated into automobiles, and some of the benefits will be realized before AVs are fully operational (Fagnant & Kockelman, 2015).

These benefits will not happen automatically. Many congestion-saving improvements depend not only on automated driving capabilities but also on cooperative abilities through vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication. Vehicle communication is likely to become standard on most vehicles even before the significant proliferation of AV capabilities, with the US NHTSA announcing its intention to mandate all new light-duty vehicles to be equipped with V2X capabilities (National Highway Traffic Safety Administration, 2014). Even without V2X communication, significant congestion reduction could occur if the safety benefits alone are realized. The US Federal Highway Administration (FHWA) estimates that 25% of congestion is attributable to traffic incidents, around half of which are crashes (Federal Highway Administration, 2005).

Multiple studies have investigated the potential for AVs to reduce congestion under differing scenarios. Under various levels of AV adoption, congestion savings due to ACC measures and traffic monitoring systems could smooth traffic flows by minimizing accelerations and braking in freeway traffic. This could increase fuel economy and congested traffic speeds by 23%-39% and 8%-13%, respectively, for all vehicles in the freeway travel stream, depending on V2V communication and how traffic smoothing algorithms are implemented (Atiyeh, 2012). If vehicles are enabled to travel closer together, the system’s fuel and congestion savings rise further, and some expect a significant increase in highway capacity on existing lanes (Tientrakool, 2011). Shladover and colleagues estimate cooperative ACC (CACC) deployed at 10%, 50%, and 90% market penetration levels will increase lanes’ effective capacities by around 1%, 21%, and 80%, respectively (Shladover, Su, & Lu, 2012). Gap reductions, coupled with near-constant velocities, will produce more reliable travel times—an important factor in trip generation, timing, and routing decisions. Similarly, shorter headways between vehicles at traffic signals (and shorter startup times) mean more AVs could more effectively utilize green time at signals, considerably improving intersection capacities (Fagnant & Kockelman, 2015).

Over the long term, new paradigms for signal control, such as autonomous intersection management, could use AVs’ powerful capabilities. Some evidence shows advanced systems could nearly eliminate intersection delays while reducing fuel consumption, though this concept is only theoretical and certainly a long way off. To implement such technologies, Dresner and Stone estimate a 95% or more AV market penetration may be required, leaving many years before deployment (Dresner & Stone, 2008).

Of course, many such benefits may not be realized until high numbers of AVs are present on the roads. For example, if 10% of all vehicles on a given freeway segment are AVs, there will likely be an AV in every lane at regular spacing during congested times, and this could smooth traffic for all travelers (Bose & Ioannou, 2003). However, if just 1 out of 200 vehicles is an AV, the impact would be nonexistent or greatly lessened. If one AV is following another, the following AV can reduce the gap between the two vehicles, increasing effective roadway capacity. This efficiency benefit is also contingent on higher AV shares. Technical and implementation challenges must be met to realize the full potential of high usage, including the implementation of cloud-based systems and city or region-wide coordinated vehiclerouting paradigms and protocols. Finally, while AVs have the potential to increase roadway capacity with higher market penetration, the induced demand resulting from more automobile use might require additional capacity needs (Fagnant & Kockelman, 2015). Travel Behavior Impacts

The safety and congestion-reducing impacts of AVs have the potential to create significant changes in travel behavior. For example, AVs may provide mobility for those too young to drive, the elderly, and the disabled, thus generating new roadway capacity demands. Parking patterns could change as AVs self-park in less expensive areas. Car- and ride-sharing programs could expand, as AVs serve multiple persons on demand.

Most of these ideas point towards more VMT and automobile-oriented development, though perhaps with fewer vehicles and parking spaces. Added VMT may bring other problems related to high automobile use, such as increased emissions, greater gasoline consumption and oil dependence, and higher obesity rates (Fagnant & Kockelman, 2015).

As of July 2014, state legislation in California, Florida, Michigan, Nevada, and Washington, D.C. mandated that all drivers pursuing AV testing on public roadways be licensed and prepared to take over vehicle operation, if required. As AV experience increases, this requirement could be relaxed, and AVs may be permitted to legally chauffeur children and persons who otherwise would be unable to safely drive. Such mobility may be increasingly beneficial, as the US population ages, with 40 million Americans presently over the age of 65 and this demographic growing at a 50% faster rate than the nation’s overall population (US Census Bureau, 2011). Wood observes that many drivers attempt to cope with physical limitations through self-regulation, avoiding heavy traffic, unfamiliar roads, night-time driving, and poor weather, while others stop driving altogether (Wood, 2002). AVs could facilitate personal independence and mobility, while enhancing safety, thus increasing the demand for automobile travel (Fagnant & Kockelman, 2015).

With increased mobility among the elderly and others, as well as lowered travel effort and congestion delays, the United States can expect VMT increases, along with associated congestion, emissions, and crash rates, unless demand management strategies are thoughtfully implemented (Kockelman & Kalmanje, 2006; Litman, 2013).

However, AV benefits could exceed the negative impacts of added VMT. For example, if VMT were to double, a reduction in crash rates per mile traveled by 90% yields a reduction in the total number of crashes and their associated injuries and traffic delays by 80%. Likewise, unless new travel from AV use is significantly underestimated, the existing infrastructure capacity on roadways should be adequate to accommodate the new/induced demand, thanks to AVs’ congestion-mitigating features, like traffic smoothing algorithms (Atiyeh, 2012) and effective capacity increases through CACC (Shladover, Su, & Lu, 2012), as well as public infrastructure investments, like V2I communication systems with traffic signals (KPMG & CAR, 2012), designed to support these capabilities. However, other negative impacts, such as sprawl, emissions, and health concerns, may not be readily mitigated (Fagnant & Kockelman, 2015).

It is possible that already congested traffic patterns and other roadway infrastructure will be negatively affected, because of increased trip making. Indeed, Smith argues, “Highways may carry significantly more vehicles, but average delay during the peak period may not decrease appreciably. Similarly, emissions per vehicle mile traveled may decrease, but total emissions (throughout the day) may actually increase” (Smith, 2013). However, AVs could enable smarter routing in coordination with intelligent infrastructure, quicker reaction times, and closer spacing between vehicles to counteract increased demand.

Whether arterial congestion improves or degrades ultimately depends on how much induced VMT is realized, the relative magnitude of AV benefits, and the use of demand management strategies, such as road pricing. Emissions are predicted to fall when travel is smooth, rather than forced, with Berry saying a 20% reduction in accelerations and decelerations should lead to a 5% reduction in fuel consumption and associated emissions (Berry, 2010). Thus, while AVs may increase VMT, emissions per mile could be reduced (Fagnant & Kockelman, 2015).

Additional fuel savings may accrue through AVs’ smart parking decisions (Bullis, 2011; Shoup, 2005), helping avoid “cruising for parking.” For example, in-vehicle systems could communicate with parking infrastructure to enable driverless drop-offs and pickups. This same technology could improve and expand car sharing and dynamic ride-sharing by allowing nearby, real-time rentals on a per-minute or per-mile basis. If successful, this has great promise for program expansions, since users could simply order a vehicle online or using mobile devices, much like an on-demand taxi, to take them to their destinations. Preliminary results (Fagnant & Kockelman, 2016) for Austin, Texas, using an agent-based model for assigning vehicles around a core region indicate that each shared AV (SAV) could replace around ten privately owned or household-owned vehicles. These simulations assumed that the SAVs operated within a prescribed 12 miles by 24 miles geofence, where trip intensity is relatively high; longer trips to or from destinations outside the geofence were not considered (Fagnant & Kockelman, 2015).

As shown in Figure 1.8, even in Seattle where vehicle use is more intense than national averages (Puget Sound Regional Council, 2006), just under 11% of vehicles are “in use” throughout the day, even at peak times, though usage rises to 16% if only newer vehicles are monitored (Fagnant & Kockelman, 2015).

Vehicle use by time of day and by vehicle age (Puget Sound Regional Council, 2006)

FIGURE 1.8 Vehicle use by time of day and by vehicle age (Puget Sound Regional Council, 2006). Freight Transportation

Freight transport on and off the road will be impacted by autonomous driving. As one example, mining company Rio Tinto is already using 53 self-driving ore trucks; these trucks have driven 2.4 million miles and carried 200 million tons of materials (Rio Tinto, 2014). The same technologies that apply to autonomous cars can apply to the trucking industry, increasing fuel economy and lowering the need for truck drivers. While workers would likely still need to load and unload cargo, long-distance journeys may be made without drivers, with warehousing employees handling container contents at either end. Autonomously operated trucks may face significant resistance from labor groups, like the Teamsters, and competing industries, such as the freight railroad industry (Fagnant & Kockelman, 2015).

Additional benefits can emerge through higher fuel economies when using tightly coupled road-train platoons, thanks to reduced air resistance of shared slipstreams, not to mention lowered travel times from higher capacity networks (a result of shorter headways and less incident-prone traffic conditions). Bullis estimates that 4 m inter-truck spacings could reduce fuel consumption by 10%-15%, and road- train platoons would facilitate adaptive braking, potentially enabling further fuel savings (Bullis, 2011). Kunze and colleagues did a successful trial run using 10 m headways between multiple trucks on public German motorways (Kunze, Ramakers, Henning, & Jeschke, 2009), and a variety of autonomously platooned Volvo trucks recently logged approximately 10,000km along Spanish highways (Newcomb, 2012). However, tight vehicle spacing on roads could cause problems for other motorists trying to exit or enter highways, possibly resulting in the need for new or modified infrastructure with dedicated platoon lanes and thicker pavements to handle high truck volumes (Fagnant & Kockelman, 2015). Anticipating AV Impacts

Since AVs are only in the testing phase, it is difficult to precisely anticipate actual outcomes. Nevertheless, it can be useful to roughly estimate likely magnitudes of impact. Based on research estimates for the potential impacts discussed above, this section quantifies crash, congestion, and other impacts for the US transportation system (including changes in parking provision, VMT, and vehicle counts). To understand how AVs’ assimilation into the road network might work, multiple assumptions are needed and explained below. To further understand the impact, we assume three AV market penetration shares: 10%, 50%, and 90%. These not only represent market shares but also technological improvements over time, since it could take many years for the United States to see high penetration rates. The analysis is inherently imprecise, as it provides an order-of-magnitude estimate of the broad economic and safety impacts this technology may have.

We assume the primary benefits for AV use will include safety benefits, congestion reduction (comprised of travel time savings and fuel savings), and savings realized from reduced parking demands, particularly in areas with high parking costs. The assumptions driving these estimated impacts are discussed in this section, as are the assumptions used to estimate changes in VMT, to estimate AV technology costs, and to select an appropriate discount rate for net present value (NPV) calculations (Fagnant & Kockelman, 2015). Changes in VMT and Vehicle Ownership

VMT per AV is assumed to be 20% higher than that of non-AV at the 10% market penetration rate and 10% higher at the 90% market penetration rate. This reflects the notion that early adopters will have more pent-up demand for such vehicles than later buyers.

Fagnant and Kockelman’s preliminary agent-based simulations (Fagnant & Kockelman, 2016) underscore this idea. For the Austin, Texas, market, a fleet of SAVs serving over 56,000 trips a day was found to travel 8.7% of its mileage unoccupied (empty). This figure fell to 4.5% when ride-sharing was permitted, and minor (less than 1%) net VMT reductions were realized when demand rose by a factor of 5 and ride-sharing was permitted. Analysis of the various simulation results suggests each SAV could serve the same number of trips as 10 household-owned vehicles (if all replaced travel were to lie within a 12 mile x 24 mile geofence). While the 10'1 replacement rate may be too high for mass adoption settings, especially in locations with lots of long-distance trip making and low-density development, 10 household vehicles are assumed to be replaced here for every SAV operated (10% of the fleet), resulting in the implicit assumption that around half of all AV trips will be served by SAVs and the other half by personally owned AVs (Fagnant & Kockelman, 2015).

Additional VMT increases may be realized from induced demand, as travel costs and congestion fall. In his review of literature spanning 30years in California and across the United States, Cervero shows that the long-term (6 years or more) urban area elasticity of VMT (demand for road travel) with respect to the number of highway lane-miles supplied ranges from around 0.47 to 1.0, averaging 0.74 (Cervero, 2001). This suggests that if a region’s lane-miles increase by 1%, regional VMT may rise by 0.74% over the long term, after controlling for population, income, and other factors. If tolls and/or other traffic management policies are put in place to stem excessive demand, demand elasticity should be lower. Of course, a 0.74 elasticity value is likely high, since AVs’ capacity effects are probably uniform, rather than targeted. Many road segments in a region are not currently congested and do not exhibit latent or elastic demand. Therefore, if we assume a 0.37 elasticity, system-wide VMT may be expected to rise 26% under 90% AV market penetration assumptions (i.e., 60% freeway congestion reduction and 15% arterial congestion reduction, due to an increase in effective capacity) (Fagnant & Kockelman, 2015).

While the congestion-relieving impacts of AVs may be similar to those of adding lane-miles, they differ in another crucial respect beyond their uniform versus targeted capacity increases, as noted above. Personal values of travel time (VOTT) may also fall because of drivers’ increased productivity gains as they are freed for purposes other than driving. Gucwa attempted to estimate the joint implications of increased travel due to capacity and value of travel time changes using simulations of the San Francisco Bay Area. When increasing roadway capacity between 10% and 100%, and simultaneously reducing the VOTT from current levels to somewhere between high quality rail and half of current (in-car) values of time, his model results produced a 4%-8% increase in VMT regionwide, because of changes in destination and mode choices (Gucwa, 2014). Of course, AVs may also travel while unoccupied, and long-term housing and employment shifts may generate extra VMT (Fagnant & Kockelman, 2015).

Cervero’s (2001) framework (with halved elasticity values) and Gucwa’s (2014) simulations produced two different VMT outcomes that may represent the respective high and low ranges of reasonable VMT growth scenarios. Therefore, we select 20% and 10% increases in VMT per AV as assumptions for the 10% and 90% AV market penetration rates, respectively, reflecting reasonable estimates within these bounds. These VMT increases are expected to apply system-wide, across personally owned AVs, SAVs and AVs used for shipping and freight (Fagnant & Kockelman, 2015). Discount Rate and Technology Costs

For NPV calculations, a 10% discount rate is assumed; this is higher than the 7% rate required by the US Office of Management and Budget (OMB) for federal projects and TIGER grant applications (LaHood, 2011). We do so to reflect the uncertainty of this emerging technology. Early-introduction costs (perhaps seven years after initial rollout) at the 10% market penetration level are assumed to add $10,000 to the purchase price of a new vehicle, falling to $3,000 by the 90% market penetration share. Internal rates of return for initial costs are included at the $37,500 level, and this may be closer to the added price of AV technologies, a couple of years after they are first introduced (Fagnant & Kockelman, 2015). Safety Impacts

US crash rates for non-AVs are assumed constant, based on NHTSA’s 2011 values, and the severity distribution of all crashes remains unchanged from the present. As noted previously, over 90% of the primary factors in crashes are human errors (National Highway Traffic Safety Administration, 2008), and 40% of fatal crashes involve driver alcohol or drug use, driver distraction, and/or fatigue (National Highway Traffic Safety Administration, 2012).

Therefore, AVs may be assumed to reduce crash and injury rates by 50%, versus non-AVs at the early 10% market penetration rate (reflecting savings from eliminating the aforementioned factors, as well as fewer legal violations like running red lights), and 90% safer at the 90% market penetration rate (reflecting the near elimination of human error as a primary crash cause, thanks to improved vehicle automation technology).

Pedestrian and bicycle crashes (with motor vehicles) are assumed to enjoy half of the AV safety benefits, since just one of the two crash parties (the driver) relies on the AV technology. Similarly, motorcycles may not enjoy autonomous status for a long time (and their riders may be reluctant to relinquish control), and around half of all fatal motorcycle crashes do not involve another vehicle. Therefore, motorcycles are assumed to experience just a 25% decline in their crash rates, relative to the declines experienced by other motor vehicles.

While safety improvements will likely be greater than new safety risks, it is possible that new risks will be greater for some system users under certain circumstances, particularly at early technology stages. Lin argues that increased safety to some users at the expense of others is not necessarily a clear- cut benefit, even if net safety risks to the whole population are lower (Lin, 2013).

In the following calculations, crash costs are estimated first based on their economic consequences, using National Safety Council (2012) guidance, and then on higher comprehensive costs, as recommended by the USDOT (Trottenberg, 2011), to reflect pain and suffering and the full value of a statistical life (Fagnant & Kockelman, 2015).

1.1.79 Congestion Reduction

Shrank and Lomax’s congestion impact projections (Schrank, Eisele, & Lomax, 2012) for 2020 are used in what follows as a baseline. They assume a $17 per person-hour value of travel time, $87 per truck-hour value of travel time, and state-wide average gas prices in 2010. They estimate 40% of the nation’s roadway congestion occurs on freeway facilities (with the remainder on other streets). By 2020, US travelers will experience around 8.4 billion hours of delay while wasting 4.5 billion gallons’ of fuel (due to congestion), for an annual economic cost of $199 billion (Fagnant & Kockelman, 2015).

This analysis assumes AVs are equipped with CACC and traffic-flow-smoothing capabilities. At the 10% AV market penetration level, freeway congestion delays for all vehicles are estimated to fall 15%, mostly due to smoothed flow and bottleneck reductions. This is lower than Atiyeh suggests so that induced travel, though additional congestion benefits, may be realized (fewer crashes, a small degree of increased capacity from CACC, and smarter vehicle routings) (Atiyeh, 2012). At the 50% market penetration level, a cloud-based system is assumed to be active; Atiyeh suggests 39% congestion improvements from smoothed flow. Further capacity enhancements of 20% may also be realized. With crashes falling because of safety improvements, another 4.5% in congestion reduction may be obtained. Again, induced travel will counteract some of these benefits, and a 35% delay reduction on freeways is estimated in this analysis. Finally, at the 90% level, freeway congestion is assumed to fall by 60%, with the near doubling of roadway capacity (Shladover, Su, & Lu, 2012) and dramatic crash reductions. However, capacity and delay are not linearly related, and congestion abatement may be even greater than these predictions at 90% market penetration (Fagnant & Kockelman, 2015).

At the arterial-roadway level, congestion is assumed to realize fewer benefits from AVs (without near- complete market penetration and automated intersection management (Dresner & Stone, 2008), as delays emerge largely from conflicting turning movements, pedestrians, and other transportation features that AV technologies cannot address so easily. Therefore, arterial congestion benefits are assumed to be just 5% at the 10% market-penetration level, 10% at 50% penetration, and 15% at 90% penetration. AV fuel efficiency benefits are assumed to begin at 13%, increasing to 25% with 90% market penetration, because of better route choices, less congestion, road-train drag reductions (from drafting), and more optimal drive cycles. Non-AVs on freeways are assumed to experience 8% fuel economy benefits during congested times of day under a 10% market penetration, and 13% at the 50% and 90% penetration levels. For simplicity, this analysis assumes all induced travel’s added fuel consumption will be fully offset by AVs’ fuel savings during non-congested times of day (Fagnant & Kockelman, 2015). Parking

Parking savings comprise the final monetized component. Litman estimates that comprehensive (land, construction, maintenance and operation) annual parking costs are roughly $3,300 to $5,600 per parking space in central business districts (CBDs), $1,400 to $3,700 per parking space in other central/urban areas, and $680 to $2,400 per space in suburban locations (Litman, 2012). Simply moving a parking space outside the CBD may save nearly $2,000 in annualized costs, while moving one to a suburban location may save another $1,000. In addition, fewer overall spaces should be needed, thanks to car sharing. Therefore, while not every AV will result in a moved or eliminated parking space, $250 in parking savings will be realized per new AV, following the earlier assumption of 10% of AVs being publicly shared (Fagnant & Kockelman, 2015). Summary of Economic Impacts

Table 1.2 summarizes all these estimated impacts. It suggests economic benefits will reach $196 billion ($442 billion, comprehensive) with a 90% AV market penetration rate. Meaningful congestion benefits are estimated to accrue to all travelers early on, while the magnitude of crash benefits grows over time (and accrues largely to AV owners/users). For example, congestion savings represent 66% of benefits, and crash savings represent 21% of benefits—at the 10% market penetration level, versus 31% and 54% of benefits, respectively, at the 90% penetration rate. When comprehensive crash costs are included, overall crash savings jump by more than a factor of three (Fagnant & Kockelman, 2015).

These results are consistent with the findings of Manyika, Chui, Bughin, Dobbs, Bisson, and Marrs (2013). These authors estimate global AV impacts of $200 billion to $1.9 trillion by 2025, assuming 5%-20% of all driving is either autonomous or semiautonomous and valuing the lowered burdens of in- vehicle travel time, at least for drivers, who can now perform other activities in route. If the 10% market penetration estimates used here are scaled globally (at least within the developed world), and the lowered burden of in-vehicle time is added, overall economic benefits are likely to fall to within Manyika et al.’s range (Fagnant & Kockelman, 2015).

Additional monetized congestion benefits may be realized beyond the values shown in Table 1.2, with falling VOTT. For example, an hour stuck driving in traffic may be perceived as more onerous than an hour spent being driven by an AV (Fagnant & Kockelman, 2015). Privately Realized Benefits

While Table 1.2 illuminates AVs’ social benefits, it is also important to anticipate the privately realized benefits of AV ownership and use. These benefits are assessed using the assumptions in Table 1.2 at the 10% market penetration and a $10,000 added purchase price, taking into account monetary savings from


Estimates of Annual Economic Benefits from AVs in the United States (Fagnant & Kockelman, 2015)

Assumed Market Shares




Crash Cost Savings from AVs

Lives saved (per year)




Fewer crashes




Economic cost savings

$5.5 В

$48.8 В

$109.7 В

Comprehensive cost savings


$158.1 В

$355.4 В

Economic cost savings per AV




Comprehensive cost savings per AV




Congestion Benefits

Travel time savings (M hours)




Fuel savings (M gallons)




Total savings




Savings per AV




Other AV Impacts

Parking savings




Savings per AV




VMT increase




Change in total # vehicles




Annual savings: Economic costs only

$25.5 В

$102.2 В

$201.4 В

Annual savings: Comprehensive costs

$37.7 В

$211.5 В

$447.1 В

Annual savings per AV: Economic costs only




Annual savings per AV: Comprehensive costs




NPV of AV benefits minus added purchase price: Economic costs only




NPV of AV benefits minus added purchase price: Comprehensive costs





Number of AVs operating in the United States

12.0 M

45.1 М

65.1 М

Crash reduction fraction per AV




Freeway congestion benefit (delay reduction)




Arterial congestion benefit




Fuel savings




Non-AV foil owing vehicle fuel efficiency benefit (freeway)




VMT increase per AV




% of AVs shared across users




Added purchase price for AV capabilities




Discount rate




Vehicle lifetime (years)




reduced fuel use and insurance, along with several levels of daily parking savings and (hourly) travel time savings.

Privately realized benefits are estimated using assumptions of a $10,000 purchase price provided in Table 1.2. These are first compared to 50% insurance cost savings from a base of $1,000 per year and 13% fuel savings from a base of $2,400 per year (American Automobile Association, 2012) over a 15-year vehicle life. Parking costs of $250 are added, representing about $1 per work day. Finally, driven time under autonomous operation is added under $1 per hour and $5 per hour assumptions, with total annual vehicle hours traveled estimated based on US average VMT (10,600 miles per year) divided by an assumed


AV Owners’ Privately Realized Internal Rates of Return (From 0% to 10% Market Share) (Fagnant & Kockelman, 2015)






Benefits (Daily Parking and Hourly Value of Travel Time Savings) (%)

$0 and $0

$0 and $1

$1 and $1

$5 and $1

$1 and


$5 and $5

$5 and $10

$10 and $10











Initial price










Mass production

$ 1 Ok









average speed of 30 mph (Federal Highway Administration, 2013). Privately realized internal rates of return are also compared to a higher added-technology price, $37,500 (Fagnant & Kockelman, 2015).

This results in the range of benefits shown in Table 1.3, across various purchase prices, values of time, and parking costs. At current high-technology costs of $100,000 or more, benefits are mostly small compared to purchase prices, except for individuals with very high values of time. Once prices come down to $37,500, persons with high VOTT and/or parking costs may find the technology a worthwhile investment. Only at the $10,000 added price does the technology become a realistic investment for many, with even the $1 per hour time value savings and $1 daily parking cost savings generating an 11% rate of return for AV owners (Fagnant & Kockelman, 2015).

We are not attempting to quantify or monetize several other possible impacts. For example, many of the nation’s 240,000 taxi drivers and 1.6 million truck drivers (Bureau of Labor Statistics, 2012) could be displaced by AV technologies, while greenhouse gas emissions, infrastructure needs, and rates of walking may fall or rise, depending on the induced VMT. Increased sprawl or automobile-style development could also result, as projected by Laberteaux (2014). Such impacts are not included in the analysis (Fagnant & Kockelman, 2015).

While exact magnitudes of all impacts remain uncertain, this analysis illustrates the potential for AVs to deliver substantial benefits to many, if not all, Americans, thanks to sizable safety and congestion savings. Even at 10% market penetration, this technology has the potential to save over 1,000 lives per year and offer tens of billions of dollars in economic gains, once added vehicle costs and possible roadside hardware and system administration costs are covered (Fagnant & Kockelman, 2015).

< Prev   CONTENTS   Source   Next >