UC Berkeley

Today’s automotive industry is witnessing unprecedented technological change. Automation is particularly expected to revolutionize how we travel and to have profound impacts on the transportation system. Whether autonomous vehicles (AVs) will improve our lives, exacerbate existing mobility challenges, or lead to currently unimagined ramifications, however, is still an open question. On one hand, the improvements in safety, efficiency, and accessibility are thought by many to be the answer to our transportation problems. However, others project a dystopian future where the efficiency improvements, while real, are not enough to counteract the trends of increasing population, urbanization, and vehicle miles traveled (VMT) per capita, as well as induced demand. While it is not certain which future beckons, there is certainty that human travel behavior, the focus of this dissertation, will be central to determining the outcome.

The objectives of this dissertation were to:

1) Collect new data on the travel behavior implications of privately owned autonomous vehicles through an innovative method that overcomes the limitations of the current literature.

2) Analyze the data to quantify the implications of privately owned autonomous vehicles on human travel behavior and the heterogeneity in the response to the technology by different demographic and lifestyle groups.

3) Integrate privately owned AVs into an activity-based model framework by estimating short-term travel demand models and proposing additional components unique to privately owned AVs.

The first step to achieving our objectives was to conduct a literature review. We identified 78 published studies that address issues of travel behavior implications of AVs. We summarized the methods currently being used to address research questions on travel behavior changes caused by AVs, highlighted their strengths and limitations, and proposed ways to improve upon these methods. We then identified critical research questions to be addressed and summarized results from the studies that addressed them. We organized the research questions into four categories: the first are research questions that have been explored by many studies, where the direction of the impact is consistent across the literature, albeit the magnitude varies considerably. For example, the impact of (shared) AVs on VMT has been well explored and most studies predict an increase which is projected to range from 1% to 90%. Similarly, many studies explored future mode choice preferences with results indicating that, overall, people prefer owning AVs over sharing them. The second category of research questions are ones with limited and consistent results, albeit the range varies widely. For example, a few stated preference survey studies indicate that reduced stress and multitasking during travel will reduce the value of time of AV riders by 5% to 55%. The third category of research questions are ones with few but conflicting results. For instance, a few survey studies indicate that people (up to 80%) do not believe their residential location will be affected by the adoption of AVs. Some simulation studies, however, indicate that lower travel costs will encourage people to move away from cities and into suburbs while other studies report the opposite. The final category of research questions are ones that received little to no attention in the literature. For instance, very few studies focus on exploring how AV owners plan to use zero-occupancy vehicles (e.g., to run errands) in order to quantify their impact on travel behavior and the transportation system.

From the literature review, we found that the two most common methods currently used to study the travel behavior implications of AVs are surveys and simulation studies. In this dissertation, however, rather than relying on surveys or simulations, we proposed a different method to explore the impacts of AVs on travel behavior: an experiment in which we simulate the experience of owning personal AVs by providing subjects with personal chauffeurs. Thus, we essentially installed the driverless feature onto their own vehicles. Just like an AV, the chauffeur took over driving duties so that subjects could relax or use their travel time productively. Subjects were also able to send out their chauffeurs to run errands that AVs will run in the future (e.g., filling up gas, picking up groceries, picking up friends and family). Subjects were tracked and their travel diaries were recorded for three to four weeks, with the outer, non-chauffeur weeks serving as control weeks (i.e., status quo conditions), and the middle chauffeur week(s) serving as treatment week(s) (i.e., “AV” weeks). By comparing travel behavior during the chauffeur weeks to the non-chauffeur weeks, we gained novel insights into what the potential shifts in travel behavior might be in an AV future.

We ran two iterations of our experiment. The first was a pilot that involved 13 subjects from the San Francisco Bay Area during the summer of 2017. The sample was a convenience sample stratified mainly by demographic (families, retirees, and millennials), where all subjects received one chauffeur week. We then ran a second, larger experiment in 2019/20 on 43 households in the Sacramento area, incorporating several improvements over the pilot study. To obtain a more diverse sample in terms of demographics, modal preferences, and mobility barriers, we partnered with the local metropolitan planning agency, the Sacramento Area Council of Governments (SACOG), who gave us access to travel survey data for a representative sample of households. We also provided a portion of our households with an extended chauffeur period (two weeks) to explore the impact of the treatment period on the results. Finally, we tracked all members and vehicles in the household and used a different phone tracking app to record a richer dataset that includes more detail on trip purpose, modes (private vs shared), parking, and vehicle occupancy.

We present two types of results in this dissertation. The first set of results are descriptive statistics that analyze basic shifts in travel behavior and the heterogeneity in the response to the chauffeur service by different demographic groups. These results are largely consistent across both iterations of the experiment. The second type of results are based on the estimation of typical travel demand models where we explore the factors behind the behavioral shifts observed in the first set of results, as well as investigate how AVs should be incorporated into an activity based-model framework. These modeling results are exclusive to the second iteration of the experiment since they require a larger sample and detailed trip data only available from the newer tracking app. Since descriptive statistics are consistent across experiments, and the second iteration included a more comprehensive set of results, we only present key findings from the larger experiment here. The sample for the second iteration was fairly representative of the population of Sacramento (the study region), albeit included a higher share of females and was more affluent and educated than the general Sacramento population. Due to the relatively small sample size and potential self-selection issues, results reported here are not projected to the general population and are only representative of our sample.

Overall, households used their household vehicles substantially more during the chauffeur weeks compared to the non-chauffeur weeks. The total vehicle miles traveled (VMT) of our sample increased by 60%, which falls in the higher end of the range reported in the literature (1% - 79%). The elderly and individuals with mobility barriers exhibited the highest percent increase in their VMT (150%) while families with kids observed the lowest increase (17%). Moreover, almost all households (95%), at some point during their chauffeur week, sent their chauffeur out alone to run errands (equivalent to zero occupancy vehicle (ZOV) or “ghost” trips in an AV future), and this made up half of the increase in VMT. During the chauffeur weeks, the overall systemwide trips increased by 25%, which drops to only a 3% increase if ZOV trips are excluded from the analysis. Moreover, subjects’ average trip length increased by 16% during the chauffeur weeks, which falls in the lower end of the range predicted in the literature (2.5% - 45%). During the chauffeur weeks, we observed a 20% increase in night trips (after 7 pm), 76% increase in trips between 20 and 50 miles, and 81% increase in trips longer than 50 miles. However, if only person trips are considered (i.e., ZOV trips are excluded), these numbers drop to 5%, 50% and 61% respectively. During the chauffeur weeks, subjects also became more auto-oriented, relying more on their “AV” and shifting away from transit trips which dropped by 70% (compared to the 9% - 70% decrease predicted in the literature). Similarly, subjects shifted away from active modes of transportation with biking and walking trips dropping by 37% and 13% respectively. The increase in vehicle miles traveled, therefore, came from three sources: 1) 50% of the increase came from subjects sending out their chauffeurs to run errands and serve friends and family; 2) 40% came from the increase in the average trip length as subjects traveled to farther locations; and 3) 10% came from subjects switching from non-auto modes to using their “AV.”

For the second set of results, we explored how to integrate AVs into activity-based models, including model specifications and parameter estimates. We investigated four components of activity-based models: activity patterns, destination choice, mode choice, and time of day, which also provided insights on the potential factors that led to the behavioral changes described above. Our formulations were inspired by the Sacramento regional model, albeit kept parsimonious with limited heterogeneity due to the small sample size. We compared the models estimated with data from the chauffeur weeks to those during the non-chauffeur weeks. We found that there were no statistically significant differences in the parameters of the individual activity patterns, destination choice, or time of day models. For the mode choice model, however, while the constant for auto did not change, the value of time dropped by 60% during the chauffeur weeks. Moreover, as the destination choice model included a logsum from the mode choice model, this resulted in longer average tour lengths, even though the parameters (beyond those in the logsum) of the destination choice model did not change. Moreover, while the trip-making propensity of individuals did not change significantly, there was a 25% increase in systemwide trip rates due to the “AV” (chauffeur) being sent on errands. This pointed to the importance of incorporating zero-occupancy vehicle trips into the activity-based modeling framework. By observing how subjects used their ZOV trips (i.e., sending their chauffeurs to run errands) we were able to propose a way to integrate these trips within a standard activity-based model framework. Our findings suggested that if ZOV trips are compartmentalized and separated from individual person trips/tours, the existing structure and parameters of an activity-based model do not need to be modified, apart from the reduction in the VOT for the auto mode. Zero-occupancy vehicle trips can then be added either as additional ZOV home-based tours or as ZOV sub-tours within the standard activity-based model process. Lastly, as inter-regional travel (e.g., tours outside the Sacramento area in our study) is modeled outside the activity-based model framework, our results indicated that modifications should be made to account for the increase in inter-regional tours, which were 54% more frequent in our sample during the chauffeur weeks.

To summarize, in this dissertation we designed and executed a unique revealed preference AV experiment that allowed us to quantify many of the potential travel behavior changes that might result from AVs. A key benefit of having access to an “AV” was the enhanced mobility and accessibility our subjects experienced during the chauffeur weeks, which was manifested by the increase in average trip and tour lengths and was highlighted by many subjects, e.g., “I love the chauffeur service. I’ve already gone to two places I would never have driven to on my own and it’s been wonderful.” At the other end of the spectrum, however, an undesirable consequence of private AV adoption was the increase in car usage which led to an increase in overall VMT and a shift away from transit and active modes of travel. Mode choice and destination choice model estimations indicated that the primary factor behind these behavioral shifts was the reduction in subjects’ VOT for the car mode, leading to an increase in accessibility (as measured, for example, via the logsum). The experiment also highlighted another undesirable consequence of private AV adoption, which was the reliance on zero-occupancy vehicles (“ghost” trips). We identified these trips as a primary source of travel behavior change, highlighting the importance of incorporating them into simulation studies. We then suggested a way to incorporate ZOV trips into an ABM framework as additional model components that consist of ZOV home-based tours and ZOV subtours using a standard ABM process. Finally, even though our sample size was relatively small, we were able to quantify the heterogeneity in the response to AVs. Results indicated that changes in travel behavior were largest for individuals with mobility barriers, the elderly, and single occupancy households and lowest for families with kids. Similarly, non-auto dependent households also observed a substantial shift in travel behavior as they became more auto-oriented.

While our dataset is for a relatively small number of individuals, we were able to obtain detailed revealed preference insight for each of these individuals into their travel behavior choices with privately owned AVs. To our knowledge, this is the first such exercise using this chauffeur approach, and we were able to quantify important travel behavior metrics for privately owned AVs as well as estimate traditional (albeit parsimonious) travel demand models. Our results provide quantitative and qualitative information on both the many benefits of privately owned AVs (“the beauty”), but also the potential drawbacks of their adoption (“the beast”). As policy makers contemplate regulations for AV deployment, it will be critical to identify and evaluate the tradeoffs between enhancing the quality of life versus the environmental and social costs of the additional travel induced by AVs.