Self-Driving Vehicles and Transportation Markets Anton J. Kleywegt School of Industrial and Systems Engineering Georgia Institute of Technology 4 September 2018 1 / 22
Outline 1 Introduction 2 Vehicles 3 Software 4 Infrastructure 5 Markets 6 Services 2 / 22
What Motivated Me to Work at Uber This Past Year? A typical person may experience 2 or 3 societal revolutions in his/her lifetime, that change the way people live and work For example, I think the following societal revolutions have taken place in my life so far: A computer on every desk : ubiquitous affordable computing devices The internet and the world wide web: ubiquitous affordable information (good and bad) I think we are experiencing the early stages of a transportation revolution: ubiquitous affordable transportation I want to be involved in that revolution I want to tell my child, maybe grandchildren, that I helped to make it happen 3 / 22
Ride-hailing Services Ride-hailing services such as Uber, Lyft, and Didi have introduced important innovations, including hailing a ride and paying for it through an easy-to-use smart phone app, and enabling a large number of people to participate in the market (as riders and drivers) This is only the beginning the larger changes are yet to come 4 / 22
Self-driving vehicles will affect various industries Vehicle design and manufacturing Software for self-driving vehicles Infrastructure Transportation markets Service industries 5 / 22
Impact of self-driving vehicles on vehicle design and manufacturing The design of both passenger vehicles and freight vehicles will be affected, possibly quite dramatically The design of passenger vehicles: Most passenger vehicles on the road, especially in developed countries, are sedans or SUVs or family vans, that is, vehicles that can carry 4 6 people at a time Most passenger vehicles on the road, especially in developed countries, carry 1 person at a time Most passenger vehicles on the road, especially in developed countries, are owned by households When a household buys a car, typically they want the car to be sufficiently versatile to be able to carry a 4 6 people, even if most trips involve only 1 person 6 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Self-driving vehicles are much more expensive than human driven vehicles, and will probably remain so until, maybe, human driven vehicles go the way of horse-drawn carts Most self-driving passenger vehicles will be purchased by (large) fleet owners, and not households An optimal fleet portfolio for a fleet owner is not necessarily a homogeneous fleet of vehicles, such as Waymo s fleet of Chrysler Pacifica minivans 7 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) If most trips continue to involve 1 or 2 persons in the car at a time (and there is reason to believe that it will), then it makes sense for many of the cars in the fleet to be designed to carry 1 or 2 persons at a time (even if such cars would not be desired by most households for a family car) The physics of shared rides implies that a very high demand rate is needed before the cost savings dominates the value of people s lost time Caveat: As long as the greatest part of the cost of a self-driving car is the cost of the sensors, cameras, LIDAR, etc, the incentive to own such non-traditional 1 or 2 person cars is not very strong, but as the cost of the sensors decrease, the incentive to own, and manufacture, non-traditional cars will increase As the numbers of 1 or 2 person cars increase, the importance of shared rides in a ride-hailing portfolio may decrease (but maybe not shared rides help to rebalance cars when demand is imbalanced) 8 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Economics of large passenger vehicles, such as trains and buses: If the cost of the driver s time is relatively large (typical in developed countries), then it is more economical to have larger transit vehicles, such as longer passenger trains (typical MARTA trains) and larger buses, because that decreases the cost per rider-trip Without a driver, the economic advantage of larger transit vehicles disappears 9 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Economics of large passenger vehicles, such as trains and buses (continued): Vehicles such as trains with a dedicated right-of-way still have capacity advantages a large number of people can be moved per unit time with no significant congestion delays Self-driving trains will operate with greater frequency, for example, one train every 2 minutes versus one train every 20 minutes (and as a corollary, each train will be much shorter than current trains, for example, 1 or 2 cars versus 10 cars) Greater frequency makes train travel more attractive, and thus the popularity of rail transit may increase with self-driving vehicles Self-driving cars may assist this greater use of transit by providing last mile transportation moving people between homes and train stations, or between work/school and train stations With smaller transit vehicles and more regular service, the separation between transit and taxi will decrease 10 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Economics of large passenger vehicles applies to large freight vehicles as well: If the cost of the driver s time is relatively large (typical in developed countries), then it is more economical to have larger trucks Without a driver, the economic advantage of larger trucks disappears With self-driving trucks, we may see a greater variety of smaller trucks, even on the interstate roads, and especially for pickup and delivery in the cities 11 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Connected vehicles Motivation: Connected vehicles can move safely with smaller headways between vehicles, thereby increasing the capacity of roads The major benefit is likely to occur on congested freeways On street networks with at-grade intersections, most of the time lost occurs waiting, decelerating, and accelerating at intersections, and thus decrease in headways is likely to provide relatively little benefit On uncongested freeways, decrease in headways is also likely to provide relatively little benefit Additional benefits require coordination with infrastructure (more about that later), to manage flows of platoons of vehicles 12 / 22
Impact of self-driving vehicles on vehicle design and manufacturing (continued) Soft connection via communication, versus hard physical connection Communication between vehicles much easier to implement Soft connection with communication results in relatively little reduction in headways, because allowance has to be made for vehicles with lowest rate of deceleration With hard physical connection, headways can become much smaller, and differences in braking abilities matter less because physically connected vehicles help each other brake, but attention has to be given to fair allocation of energy consumption 13 / 22
Software for self-driving vehicles Just as manufacturers of computer hardware do not have to be (and are not) dominant developers of computer software, so it follows that manufacturers of self-driving vehicle hardware do not have to be dominant developers of self-driving vehicle software Similarly, just as developers of computer operating systems (such as Microsoft) and online markets (such as Amazon) are different, so it follows that developers of self-driving vehicle software and ride-hailing markets are likely to be different too Just as in most computer software application areas, such as computer operating systems, there often are many initial contenders but only a few dominant ones survive, there are many contenders for software to control self-driving cars, and only a few dominant ones will survive 14 / 22
Software for self-driving vehicles (continued) There are roughly two approaches to software for controlling self-driving vehicles: Model based approach: Traditional approach to control systems develop a model of the physics of the system and its environment, and then develop controllers based on the model that collects data and selects controls to pursue an objective Black box approach: Select a collection of basis functions (such as a neural network architecture), collect a large amount of input/output data, and select the parameters (training/learning) of the architecture to map input to output 15 / 22
Software for self-driving vehicles (continued) Early software for self-driving vehicles based on the model based approach It was found that with modest engineering effort (and a very large data collection effort), a (somewhat surprisingly) good self-driving controller can be developed using the black box approach The black box has some serious shortcomings though: The black box controller makes mistakes (these are euphemistically called edge cases ) too frequently to be acceptable for widespread use in self-driving vehicles Whereas the model based approach is interpretable, the black box approach is not interpretable (this can be used as the defining separation between the two approaches) As a result, when the controller makes mistakes, if the model based approach is used, then usually one has a good idea what to fix, but if the black box approach is used, one does not have a good idea what to fix 16 / 22
Software for self-driving vehicles (continued) The black box has some serious shortcomings (continued): The typical remedy when using the black box approach is to collect more training data in the setting in which the mistake was made This remedy may fix the considered mistake, but often it opens up new mistakes in different settings the behavior reminds me of the game Whac-A-Mole, especially the version in which the moles are connected so that when one mole is whacked, another automatically jumps up Much of the data collection for current black box controllers is environment specific (and it is very expensive), so that a large data collection effort is needed when the operation of self-driving vehicles is extended to a new or larger network, that is, the black box approach does not scale well The dominant controllers for self-driving vehicles are likely to use the model based approach 17 / 22
Self-driving vehicles and infrastructure All the well-known developers of self-driving vehicles claim that they are developing self-driving vehicles to operate autonomously, that is, without the need to communicate with other vehicles or with infrastructure Infrastructure can make the system safer and more efficient: Self-driving vehicles struggle to read traffic signs and see traffic lights in rain and snow, and self-driving vehicles cannot observe traffic beyond their neighboring vehicles Infrastructure could collect and communicate needed information more reliably The autonomous vehicle approach requires every self-driving vehicle to have a full set of sensors/cameras/lidar and other devices, which is very expensive, has to be moved everywhere the vehicle goes (thus requires greater expenditure of energy), and is likely to be damaged when the vehicle is in an accident Infrastructure could collect and communicate needed information, does not have to be moved, and can be given more robust protection 18 / 22
Self-driving vehicles and infrastructure (continued) Infrastructure can make the system safer and more efficient (continued): Controllers for self-driving vehicles, especially based on the black box approach, exhibit bad equilibria: Competing controllers are being developed by collecting data while driving in traffic It has been observed that when other (human) drivers drive more aggressively in the presence of self-driving vehicles (collecting data in training), then the self-driving vehicles drive more defensively and thereby learn to drive less aggressively, and vice versa Developers of controllers have an incentive to train their controllers to drive more aggressively, and to force their competing controllers to drive more defensively If infrastructure controlled the movement of self-driving vehicles, then such bad equilibria can be avoided, and the traffic flow can be safer, more fair, and more efficient 19 / 22
Self-driving vehicles and infrastructure (continued) Infrastructure can make the system safer and more efficient (continued): Who should provide and operate infrastructure? Infrastructure for self-driving cars should be provided and operated by an organization separate from the fleet owners/marketmakers, and should be shared by all self-driving vehicles The same infrastructure used to control the movement of self-driving vehicles, can be used to collect user fees Shared infrastructure is likely to be important for the control of self-driving vehicles, and road user fees will become much more widespread 20 / 22
Self-driving vehicles and transportation markets With human driven vehicles, we can distinguish the following markets: Market for longer term vehicle rentals, typically in multiples of 24 hour time periods Market for shorter term vehicle rentals, sometimes called car sharing, typically by the minute or hour Market for rides in contrast with vehicle rentals, the service includes a driver There are significant economies of scale with such services the greater the density of vehicles, the shorter customers average waiting times or average travel times to an available vehicle Such economies of scale favor one marketmaker for each metropolitan area With self-driving vehicles, the separation between these markets decrease (and competition will likely increase) Thus, there are multiple paths to transportation markets with self-driving cars which is more likely to be successful? 21 / 22
Self-driving vehicles and service industries Self-driving vehicles need the following services: Refueling Cleaning Maintenance There are also significant economies of scale with such services the greater the density of facilities that provide such services, the shorter the average distance and time a vehicle has to travel to obtain such services Such economies of scale favor a network of service providers that service all self-driving vehicles (as opposed to each fleet owner having its own service facilities) Maybe such service facilities will be the future tenants of current parking facilities 22 / 22