The Once and Future AV
This article is for those who have passively consumed press about Autonomous Vehicles or Self Driving Cars over the past couple years, trying to track the changing topology of such a vast, complex landscape. This article intends to build a map around the events, technology, and possibilities that continue to shape this winding narrative.
Special thanks to Thyanna Voisine, Roger Ignazio, Heidi Eggert, and Shashwat Udit for fact and sanity checks
In the beginning, humans hunted and gathered. We lived in small roving tribes fighting the elements and each other, all while discovering fire and cooking, clothing and the wheel. Eventually settling into agrarian societies, we made one of the greatest discoveries in history: semi-autonomous transportation. These vehicles helped us farm more effectively, maneuver quickly, and better battle each other. They required minor supervision once properly trained, and were hesitant to put themselves in danger, with the happy happenstance of sparing their human passengers. On the whole, horses, oxen, and other beasts of burden were pretty good. The only downside is that these transports needed fed, and someone had to clean up their waste which, luckily, was a great fertilizer. On the backs of this technology humankind rose to master the Earth. Then, in 1885 we ruined everything by inventing the internal-combustion powered automobile. Suddenly, transportation required constant human engagement, lest the equipment run into obvious objects like trees or pedestrians. Furthermore, these new vehicles consumed processed fossilized hydrocarbons and emitted global death-gas, but they did go faster than camels. One-hundred and thirty years later, we’re on the verge of converging the speed and power of mechanical vehicles, with the self-preservation and learning capability of beasts.
The invention of autonomous vehicles (AV) capable of driving themselves from point A to point B without human intervention is the most disruptive technology of the 21st century, but it’s hardly a new idea. As automobiles proliferated and replaced horses in city traffic, people realized that driving is a tedious and dangerous activity. In 1939, General Motors sponsored a World’s Fair exhibit, called Futurama, purported to peek into the year 1959. Although in modern times we’d find the show mundane — it focused on the value of an interstate highway system and single-direction roads for high-speed travel — it did demonstrate a vision of a future with an automated highway system, purported to fix issues like city pollution and the growing death toll from automobile accidents. Today, we speak of autonomous vehicles as a future technology, but long ago it was thought to be a mere 20 years away; we’re sixty years late.
Automobiles are involved in the deaths of an estimated 1.3 million people per year, and 20-times more are injured. This figure surpasses all deaths from wars, drugs, and violent crime combined. Automobile accidents are the number-one killer of teenagers and young adults. In the time it took you to read this single sentence, five people have been injured or killed in an automobile accident. If any disease or military attacks facilitated such high death tolls, we would hold endless vigils and politicians would run on platforms calling for their eradication.
But there is good news: the problem of allowing distractible, simian-brained humans to operate vehicles is a vincible one. Mankind has a long history of removing people from dangerous or hazardous jobs and replacing them with machines. The advent of autonomous vehicles is just another case study in our ingenuity.
The average driver might clear over 800,000 miles in a lifetime and be involved in an accident once every 200,000 miles. An intelligent digital driver trained with the combined experience of an entire fleet of cars can rack up billions of miles in the real world, trillions in simulations, and participate in more rare situations than any individual in history. Every year the brains behind autonomous vehicles get smarter, and unlike every 16-year-old with a learner’s permit, they don’t need to start from scratch for each new driver. Every new accident or novel situation makes the whole system smarter. The system continuously trains an AI model and periodically deploys updates to the entire fleet. Self-driving cars commoditize transportation.
Like all emerging technologies, promises must be balanced with perils. There’s an assumption that autonomous vehicles will reduce the number of cars and trucks on the road, making travel safer. But the system can also be gamed: empty “zombie” cars might circle the block to avoid parking fees, reducing income for many cities and towns, and thus budgets for the operation of emergency services and local government. We can hope that autonomous vehicles will reduce road fatalities, but as Einstein put it, we may be underestimating the infinite potential of human stupidity. We can be certain that people will find ways to make something that is inherently safe quite dangerous in other ways, such as inventing a new trend of surfing on the hood of a self-driving car.
AVs are neither a curse nor panacea but depend on how we choose to use them. What should be clear by the end of this article is that the benefits and opportunities are immeasurable, and likely outweigh potential risks and unknown externalities. The technology is nascent, exciting, and increasingly inevitable.
Lives, and Other Benefits
To focus exclusively on autonomous human transportation vehicles misses much of the revolution that’s happening right now. Pit mining equipment is moving toward full automation. Electric vehicles are disrupting the automotive and fuel industry, and related supply chains. Ride-sharing is changing the culture of car ownership, and by extension, laws and city planning. AVs will only accelerate this change. Autonomous ground and aerial package transport drones are improving last-mile package delivery (i.e Fedex and Amazon). The physically impaired will gain the freedom of independent mobility. Autonomous vehicles will save lives. AV makes way for endless, intoxicating possibilities.
Personal Transportation
A self-reliant Indiana farmer, my grandfather drove everywhere in a sort of rural vehicular maximalism. He operated an assortment of arcane equipment to coax crops from the capricious Hoosier soil, from “planters” to “combines”, “threshers” to tractors hauling “balers”. As a volunteer fire chief, he drove the fire engine some days, and others, the ambulance. Friends and family could always count on his help, from moving homes to construction in a shuddering Ford F-150 XL. Outside of work, my grandparents would tool around town in their Ford Taurus, which only my grandfather drove. That is, until he passed away, stranding Grandma alone in a remote farmhouse. After a lifetime of being my grandfather’s passenger, she now relies on others to drive her around. After 80 plus years, her autonomy vanished overnight. An AV would be more than convenience for her, but freedom.
But benefits of AV go beyond mere autonomy. Grandma’s on a fixed income. Whether you lease or own, there’s a monthly cost to pay for any vehicle. There are taxes, emissions testing, insurance, cleaning, fuel costs, oil changes and unforeseen costs like accidents. Picture a few community AVs maintained by a fleet company and rented out to dozens of neighbors at a low cost. All costs associated with vehicle ownership are now divided. The cost of personal transport car can shrink from an estimated $1.50 to 25¢ per mile.
As we look toward urban areas, this value proposition improves. Urban drivers travel shorter distances, and economies of scale increase availability driving prices even lower. The ability to provide mobility as a low-cost service is a need as old as urban planning. You want a mobile citizenry to spur economic activity, in terms of taking individuals to restaurants, stores, and work. Public transportation is a complicated topic, but every city is equipped to handle cars. No matter your income, an app to safely request an autonomous car from your home to a destination for a nominal fee is the best use of an individual’s time vs. cost, compared to standing at a bus stop. Think of ride-share apps like Uber or Lyft, but community owned and operated by a fleet management service.
As transportation trends from a commodity to service, businesses are starting to offer them as an incentive. The ride app Freebird helps businesses cover the cost of transporting customers to a given location. For example, you choose a bar, an Uber picks you up and takes you there. Once you’ve spent a minimal amount, say, $10, then the bar will reimburse you for the ride. Imagine transportation as an employee perk. Providing a company car is out, paying your Lyft bill is in. AV chauffeur services are cheaper than building a new campus parking garage and allows us to work (or play) during the commute.
As prices continue to dive, fleet owners will need to differentiate themselves. Las Vegas still has a robust limousine industry because people on this sort of vacation are more likely to splurge on a fun and ostentatious nightclub transport. Not all competition will be based on price. Extras like in-car massages or comfortable seats with wide-screen movies are choices that consumers will want and pay for. Having a suite at an NFL game is fun, but the experience can begin at pick-up, with pre-game announcements, and probably lots of beer ads, and beer itself. What does it mean to drink and drive when there is no driver?
City Planning, Real-estate, and Hospitality
With the rise of ubiquitous personal transportation as a service, it’s easy to imagine how traffic patterns and human behaviors will shift. Those changes will affect how cities plan huge swaths of real estate, repurpose parking lots and garages, and where humans choose to live. In many cities, around one third of all usable space is dedicated to parking. In a world where shared AVs are generally on the move, dropping people off and immediately picking up a package delivery, much of that space can be reclaimed. This new availability of valuable space could shrink the cost of real estate in city cores or grow the amount of greenspace. Useable commutes, commutes you can devote to things other than driving, will alter how long people will be willing to travel to work. People will be willing to live further away, disrupting real estate further.
As AVs become the dominant form of personal transport much of the infrastructure of traffic management can disappear. AVs are efficient drivers. That paired with IoT transmitters to communicate with other AVs at the speed of light, vehicles can negotiate their intention to turn or drive through an intersection. When the speed limit is codified in a shared digital map that all AVs can access, traffic signs and lights become unnecessary. Why bother with speeding cameras or traffic cops when cars are programmed to never go above a legal limit? Removing the need to maintain all of this infrastructure will save cities millions a year. These savings could balance the massive loss in municipal fees from parking and speeding tickets.
When roads are AV-only, the width of roads can change. Skinny lanes that cars effortlessly navigate can free up room for an additional lane, perhaps a high-occupancy vehicle (HOV) lane to support car-pooling. Or better yet, create room for more bicycle, pedestrian, retail space, or outdoor restaurant seating.
Industry 4.0
In the same way our ancestors called automobiles “horseless carriages,” calling AVs “self-driving cars” narrows our vision around their potential impact. AVs are more than Uber with a robot chaffer. The ability to move things without human intervention will change all manner of transportation tasks, from logistics to hospitality.
The first large-scale use of AVs was not human transportation, but heavy mining equipment. Rio Tinto Group built an autonomous mine in 2008 in Australia, monitored remotely, 1000k away in Perth. Caterpillar’s MineStar AV is generally available and operating globally. Another early adopter is autonomous tractors for harvesting large scale farms. Like mining, this is an easier use-case since the range is bounded. In 2015, a convergence of AI, robotics, and AV birthed Greenbot. This helpful little fellow is capable of functioning in smaller outdoor areas for a variety of general uses, from mowing lawns to picking fruit.
Like many other technologies that require the budget and support of government, unmanned ground vehicles (UGV) have existed in militaries for eighty years, since the first Russian radio controlled teletanks. Much of the modern commercial work in AVs was spurred by three million-dollar purse DARPA self-driving Grand Challenges. Most of the early commercial AV leaders such as Waymo were born of GC alumni. While the purpose of the prize was to promote research for military improvements to UGVs, it has grown into the burgeoning industry we see today. This ranks AVs as one of the more successful innovations to come from government R&D investment, alongside the Internet and Tang.
AVs also have the potential for massive disruption in Industry 4.0 supply chain operations. Starting with Amazon-style automated warehouses powered by thousands of AV picker robots, to Honeywell’s robots loading Embark’s autonomous long-haul trucks. Couple this with strides made in last-mile delivery leveraging smaller personal AV delivery systems such as Uber and drones, we’re entering an age of fully automated logistics.
Environmental Impacts
We are now firmly in the Anthropocene Era, where humans are the dominant factor affecting the global climate. We’re experiencing record carbon dioxide parts per million (CO2 PPM), routinely breaking the hottest years on record, the increased energy in the atmosphere projecting as stronger storms, bigger floods, longer droughts, rising extinctions. All signs point to rising ocean levels, displacing humans living in coastal and other newly unlivable regions. Were it not for a few quirks of history involving the interests of gilded age robber barons, there’s a chance we could have spent the last century perfecting electric vehicles (EVs) rather than internal combustion engines (ICEs). In 1917, more electric cars than gas powered ones were produced, only to be rapidly outpaced by petrol powered ICEs over the next century.
Today, the average car pumps 6 tons of carbon dioxide into the atmosphere every year, a major contributor to global climate change. An equivalent electric car, charged by connecting to a grid powered by renewable sources, creates about 2 tons of CO2. Not all electricity comes from renewable sources, so these numbers are trickier to quantify. What’s clear is that as the globe moves away from fossil fuels, EVs are becoming future, currently exemplified by the EV-only automaker, Tesla. And then there are Hydrogen fuel cell vehicles (HVs). Measuring CO2 emissions from this engine is also tricky, depending on how the hydrogen is derived. Like EVs, fuel cells don’t actually emit CO2 when they’re driven but suffer from a similar “long tailpipe” problem of spewing carbon dioxide at the moment of refinement into fuel. HV emissions plus manufacturing pollution puts them on par with current EV technology for total emissions per vehicle lifetime.
Brian Johnson, an analyst from Barclays, estimates that car sales will plummet 40% over the next 25 years as rent-per-use fleet vehicles increase. This further reduces the carbon footprint of manufacture of excess vehicles that will spend 95% of their lifetime sitting idle. Then there are the improvements from simply driving better. There is a subculture of people called hypermilers, who optimize their driving through a series of techniques like scientific route planning or drafting semi-trucks to minimize fuel consumption. Cars driven by AI can continually improve their driving habits at scale.
Taken all together, EV/HV and AVs together could reduce annual CO2 emissions from transportation by around 6.5 billion tons, a 17% reduction of global human carbon output. Only about one-fifth of that savings comes from alternative fuels, the rest from the proliferation of AVs. The unreasonable effectiveness of automation in fighting climate change in the transportation space will dwarf any particular switch from gas to EV or HV. While alternative fuels are worth pursuing for a number of reasons, if governments are truly concerned about the volume of atmospheric carbon, the best transportation investment they can make is funding the adoption of AV fleets.
Technology
In 2019, Waymo is the quintessence of AV companies, with a car that looks like a robotic marshmallow topped by an obsidian gumdrop. It’s polished and clean with utopian futurism, the kind that will eventually come to look like a dated, rusting, 1950s Electrolux. But the idea that Waymo and other AVs represent is quite old. The World’s Fair Futurama exhibit intended for driving automation to exist by 1960, through a combination of outfitting roadways and cars with electronic infrastructure. Despite the obvious benefits, AVs never took hold.
What’s Old is New Again
The missing component to make AVs work is artificial intelligence. While many technologies have converged to make AVs viable, from better sensors and fully electronic driving systems to high definition maps, the core problems faced require artificial perception. A driver can easily tell the difference between a human about to cross the road and a bus stop poster with a picture of a person on it. No simple electronic infrastructure is going to alert you to a dog running into the street versus a harmless tumbleweed. What AVs have always lacked were the kind of perception skills available to any child.
Without a robust AI able to classify objects and reliably estimate what those objects will do (“Is that man about to cross the road, or just standing there?”), AVs are dangerous murder machines, absently careening down the road blind to the world around them. In the past, attempts to make AVs absent of AI have required massive infrastructure investments, such as sensors that communicate what speed a car should move and when to stop, and lane markers that tell the car when it’s in the lane. And since these cars can’t see you, pedestrians are restricted from their driving space. The old attempts effectively envisioned AVs as tiny single-car trains, or trackless streetcars. But even if we had made such investments, AVs would never drive in your neighborhood. They would never get you home.
Autonomy
Defining what autonomy is can be tricky. Does the sensor-based infrastructure approach even count as autonomous? What about parking assist? Or lane assist, which allows drivers to take their hands briefly off the wheel and pedals on an interstate? So far, we have not been specific about what it means for a vehicle to be autonomous, implying a jump directly to full autonomy, known as level 5 by the Society of Automotive Engineering (SAE), best known for its automotive horsepower ratings. They range from level 0 (no automation) to level 5 (full automation).
As of 2019, level 5 automation is still a ways off, but major initiatives are underway in its pursuit. One side promotes the idea that full autonomy can be accomplished by walking the ladder of automation, starting with partial automation (level 2), moving up to conditional automation (level 3), slowly improving to level 4 and eventually to level 5 given enough tweaking. Tesla exemplifies this approach. There is a fair bit of research showing that when humans believe that they’re being assisted, they pay less attention to the task at hand, thus increasing the likelihood of error. This alone should dissuade the idea of consumer-wide level 3 automation as a dangerous endeavor.
Most other AV companies, led by Waymo and Cruise, are instead attempting to jump directly to level 5 autonomy. As a stopgap, AV companies are driving level 4 automobiles on the road under the supervision of trained professional drivers who take over when they encounter new situations, such as inclement weather, and restrict themselves to specific geo-fenced locations. A variant of this approach is remote monitoring, with humans ready to disengage autopilot when necessary.
Waymo has been running a fleet of level 4 AVs in Phoenix, AZ for paying customers since winter of 2018. The cars are called to a pickup location, like any ride-sharing app would, and the vehicle drives up automatically. The user enters and the car takes them to a destination. The car contains a safety technician to take over in anomalous edge cases called disengagements. Technically, the same dangers due to human inattentiveness that exist with level 3 exists for level 4, just at a less frequent pace. It’s this lack of clarity for when a car moves from level 4 to 5 that requires a persistent human figure until the incident rates drop below acceptable levels. “What’s an acceptable level,” is an open question, but we can create some reasonable goals.
The average driver is involved in an accident once every 180,000 miles. We want AVs to be better, so if we chose a benchmark of 500,000 miles an AI driver, it’s a simple matter of driving cars around with human technicians involved until a fleet averages half a million miles per disengagement. At that time, we can consider the AV to be level 5 for all intents and purposes, fit for consumer use, safer than humans. This alone should cut casualties in half. Then improvements in technology become a systemic improvement, jumping from 2x safer, to 4x, 8x, 16x, and so on. The escape velocity of safety will be reached, where no human driver could match, thus crowning AVs with safety supremacy.
AV Technologies
Modern AVs are a bubbling hodgepodge of technologies and disciplines. Breaking down all the tasks that human drivers accomplish as pilots could fill volumes. The technology required to then automate those feats, even more so. A high level short list of the things that human drivers do automatically, or with a little training is: navigate to a destination, control the car’s speed and direction, adjust based on feelings like tire skid, visually perceive the world around them, predict the likely movement of other agents to react accordingly.
For an AV, all of these inputs and outputs must be controlled by an array of various sensors and controllers fed into a set of filters and artificial intelligence decision makers. Fundamentally, all of these systems exist to help answer in real-time the pressing questions: “How fast do I go and in which direction?”. This can be broken down.
Control Systems
First, we must consider how a computer can control a car at all. All of the perceptual technology in the world won’t matter if a car can’t move. Initial attempts at AV, such as vehicles in the first DARPA Grand Challenge, required actuators to press the accelerator and brake pedals, servos wired to control the steering wheel, managed by a standalone computer system. As modern cars have become increasingly computerized, later attempts connect directly into drive systems with software, but still require a control system. One example system is the proportional-integral-derivative (PID), that acts as a continuous feedback loop that can smooth out simple functions, like cruise control. These are in addition to the necessary vehicle telematics for monitoring basic details like whether the battery needs changed or checking if the windshield wiper fluid is low.
Navigation
With the physical aspects of controls engineering solved, navigation requires different disciplines. One represents your feet and hands, the other, your ability to plan and follow a route from point A to point B. One of the fundamental needs in driving a car is localization, or basically, knowing where a car is on Earth, where it is on the road, what direction its facing, its bearings, velocity, and more. The impairment of such tools is what makes a drunk such a terrible driver. This intersection of knowledge is achieved by crossing the physical position, orientation, and velocity of the car with a high definition map of the world (containing data such as roads, speeds, intersections and stoplights). Creating high definition maps starts with detailed satellite images, augmented with more detailed maps populated by aircraft (think B52s), and provided with more detailed 360 camera maps using general purposes mapping vehicles. More up-to-date data comes from public and private traffic and construction information, alongside real-time app data from millions of drivers’ smartphones from apps such as Waze. These maps are kept live by sharing streams of sensor data from all AVs in a fleet.
There are a useful set of older technologies on board that aid in navigation, broadly called dedicated short-range communications (DSRC). This is how autos can communicate with each other, (vehicle to vehicle (V2V) or vehicle to infrastructure (V2I). DSRC is the technology originally envisioned from the Futurama days, but is still an area of active research by various departments of transportation. It has taken on a new urgency for communicating to increasingly computerized vehicles at scale. Localization and HD maps all come together to create a digital model of the world and the car’s place in it called the occupancy grid.
Inertial Navigation System Knowing the orientation of a car largely requires older technology such as a global navigation and satellite system (GNSS), and inertial measurement unit (IMU), which contain much of the same technology that exist in any smartphone. GNSS can pinpoint the car’s geo-spatial position on the planet. The most famous example is the global positioning system (GPS) created by the US military, but there are other governmental and commercial geo-spatial systems that can be used, such as Beidou and GLONASS. An AV could be designed to use any or all of these systems, so we just call it GNSS.
Beyond global position, IMUs detect the car’s pose, meaning where the nose is and its pitch, yaw, and roll. These are nautical terms originally created to describe the orientation of a boat, but much more precise. For accuracy, IMUs are a combination of technologies, leveraging GNSS, odometer, accelerometer, gyroscope, and compass data. It knows the position, orientation and velocity of the car. Think of an IMU like a human’s inner ear, keeping the car oriented and balanced, while GNSS is your knowledge of where you are in your neighborhood.
These data sources converge to plan and execute the AVs path.
Sensors and Sensor Fusion
After control, orientation, and navigation we need to add artificial perception, which requires various vision systems, stitched together in a technique called sensor fusion. While there is still some debate about how fancy a vision system needs to be, all AV work assumes that human-level cameras are necessary. Many landmarks on a road assume visual acuity. For a driver to react to a speed limit or stop sign still necessitates passing a vision test, and AVs are no different. But some electromagnetic sensors, such as LiDAR (light detection and ranging), provide super-human perception of the world around them. It is due to these extrasensory technologies that give so many industry experts the belief that AVs will eventually best any human driver on Earth.
While humans can only face one direction at a time, LiDAR can detect in 360 around the car, through occlusions like fog or dust, and work at night. AVs also leverage RADAR for forward collision warning (FCW) and avoidance of large objects, but suboptimal for detecting smaller objects like pedestrians or cyclists. This is the same technology used for parking assist sensors, the kind that beep if you’re about to back into a fire hydrant. All of this sensor data is fed into a sensor fusion system (Extended Kaplan Filters) with increasing levels of recognition, from signals, to characteristics, to symbols, with each level providing a more holistic understanding of the world. These sensor fuse to create a seamless whole from an ever-shifting picture of the state of the world around the vehicle, just like human brains do, creating a conscious narrative from glimpses of perception.
More and More and More
The technical complexity of building autonomous vehicles is to train an AI to operate confidently and correctly within a cone of uncertainty. This goes beyond aligning sensor data to move on a high definition map. It also means sensing what other objects are doing, and predict their most likely path, such as, “cars will likely stay on the road, cyclists will be off to the side, pedestrians will be on the sidewalk.” Of course, we know that each of these rules can change at any moment, so the AV also has to detect alternative possibilities. Pedestrians can change course to chase a runaway dog across a road. Then there’s the subtle rules of being a good citizen on the society of roads. Cars need to signal to humans where they’re going. Real people expect to see human drivers creep into an intersection a bit as a signal to others that they believe it’s their turn to go. Humans often wave cars on, and sometimes a traffic light is out and requires a police officer to direct traffic. For all of these conditions, and thousands more, we need to ensure an AV is trained to handle before we can confidently call it a level 5, fully autonomous vehicle.
Not There Yet
Before level 5 AVs can reach their true potential, we have a long way to go on various fronts. We need to finalize some technology needs and legal implications, and must work on finding answers to some of the more dire predictions, like what do we do when millions of professional drivers are suddenly out of the job? While AVs are one of the most disruptive technologies to appear in the history of transportation, it’s the disruptive iceberg below the waterline that account for the biggest impacts.
Educating AI
There’s a bad joke in computer science called the 90/90 rule: the first 90% of the work accounts for the first 90% of the time, while the last 10% of the work accounts for the remaining 90% of the time. Autonomous vehicles are called “AI’s greatest problem” because the results of minor mistakes can be so dire. Few artificial intelligence problems require such speed and precision where lives are directly on the line, where a model that is incorrect less than 1% of the time is still involved in accidents on a daily basis. Training an AI model from 99% accurate to 99.5% accurate could take 10x more effort, while going from 99.5% to 99.6% can take 100x more, and so on. This is the fundamental reason why in 2019 AVs seem stalled at Level 4 autonomy. Improving the AI to drive safely for 500,000 miles between crashes is orders of magnitude harder than 100,000 miles. As a vehicle stays on the road longer, the odds of encountering rarer edge-cases appear, and AI must be trained on most of them. Here are a few.
Weather
First, there’s the issue with weather. While AVs are working well in arid climates, even slightly more inclement weather drastically increases difficulty. There are simple weathers like rain, and complex phenomena like snow and ice. What should an AV do when confronting a tornado in Oklahoma? We’ll never reach a point where the AI that powers an AV makes the perfect choice in every scenario, and it’s not necessarily a bad thing to prefer some human control over the decision-making process in extreme cases, if only as an option: drive through this high water, or wait for the flood to get worse?
Moravec’s Paradox
Problems that are relatively easy for a 1-year-old to decide but difficult for a computer are captured in something called Moravec’s paradox. AVs need to anticipate the behaviors of others. Is that runner about to cross the road or just stretching at the crosswalk? Is that person in the road a police officer? If so, are those hand signals directing traffic? Does the added context of that broken traffic light matter?
Such a lack of human-level common sense expresses itself in many ways. In 2019 they are exemplified by examples of AVs confusion at recognizing stopped cars or facing difficulty making left turns. When Aptiv cars confront a line of unexpected street parked cars, it may also stop under the assumption that they are all in line to make a right turn. Or vice versa. As for making left turns, these are complicated maneuvers that even human drivers have trouble with. The GM-backed AV company Cruise has made pretty good progress here by focusing a huge effort on tackling this particular complexity, but it’s going to require continued herculean efforts for any new player in the AV space, or a new cottage industry to provide ready-to-use, pre-trained AI models.
Security
Security is always a concern with any computer system, but for a computer network controlling a fleet of one-ton terrestrial projectiles flinging people and property, the need for robust security is greater. Think military grade security, with medical grade testing. There are a million horror scenarios for criminal activity to access a vehicle, from theft of shipments via redirecting its movement, to kidnapping, to remote assassinations of ranking officials by causing wrecks. The ability to transport drugs and victims becomes much less risky. We need solutions to these real possibilities. We also require a robust infrastructure where the details concerning an accident are accessible by the right authorities.
Few people in the world are able to do the work of building autonomy, spawning a cottage industry to get folks up to speed. Udacity is an online training course educating engineers in the basics of the necessary tools from creating and leveraging artificial intelligence models to improving the millions of lines of code necessary to create a basic AV. These efforts are important and necessary, and opportunities in the fractal periphery of the AV field will continue to multiply.
Laws and Employment
Despite secret wishes of luddites everywhere, AVs are going to happen, at scale, and soon in some capacity. The economic upsides are just too great. Perhaps we’ll never see full personal driverless cars, but long-haul shipping, geofenced transportation like airports, and remotely operated consumer fleets are already happening. But before our system can support widespread adoption, there are current legal issues that need hammered out on many fronts. These will certainly be lobbied to change for the benefit of corporations with vested interests. Walmart is sometimes called a “mobile warehouse” due to the vast number of trucks on the road at any time. Along with Uber and Google aggressive government lobbing efforts are already underway. Laws won’t be the thing to stop AVs.
“Any motorist who sights a team of horses coming toward him must pull well off the road, cover his car with a blanket or canvas that blends with the countryside, and let the horses pass. If the horses appear skittish, the motorist must take his car apart, piece by piece, and hide it under the nearest bushes,” reads a late 1800s Pennsylvania law concerning the newly invented automobile. As laws are altered to adjust to the existence of AVs, we need to avoid repeating the history of absurd laws regarding new technologies.
Employment
What do we do as a society when AVs automate more and more driving jobs? With over 3M people employed in industrial transportation in the US alone, entire logistics operations are likely to be automated, causing mass unemployment in those fields. AVs will certainly create new jobs and industries we aren’t aware of yet, but in the interim many people will find their skills valueless in the labor market. As a society we need to have a plan on what to do when the skills that millions of people spent a lifetime perfecting are no longer in demand in the marketplace. As NYC Traffic Commissioner Samuel Schwartz once said, “cars first, people second, is a mindset that has been difficult to change.” Now is the time we need to change that mindset.
Another industry to be hit by AVs is automobile manufacturing. If the promise of ever-present, on-demand AV fleets come to pass, car and truck ownership will drop. While a reduction of needed cars is good news for the environment and consumers, this means that automakers will make fewer sales, necessitating drops in assembly, parts manufacturing, materials, and all related supply chains.
Drive through any small American town, and you’re likely to see a few things. Gas stations, auto mechanics, muffler shop, oil change stations. As these are consolidated due to shrinking margins and new technology such as EVs, many places will have few employment opportunities left. In addition to the lost revenues of small businesses, and the associated taxes, many municipalities rely on traffic and parking fees to balance their operational budgets. This sudden loss of cash will have all manner of effects, from how many police the city can employ, to parks and recreation cuts. AVs can be programmed to never break laws. Outside of law and employment, there are more systemic disruptions waiting in the wings.
Accidents and Death
Whatever the unknowns and disruptions, we owe it to our descendants to create them anyway. AVs’ greatest opportunity is reducing the 1.3 million annual casualties due to automobile incidents. While reduced traffic deaths is a self-evident good, there are downsides. The majority of organs available for transplant come from healthy people who died in automobile accidents. If the death toll in this way drops, the death toll for people on the organ transplant list will increase.
Or consider a case where an AV driving on an icy patch of bridge or another force majeure puts a pedestrian in the car’s path. The car can choose to either kill the pedestrian or run the car off the bridge killing the passenger. How should a computer handle this situation? How would a human? Most of us would react quite differently depending on who the pedestrian was. We might be more willing to strike a single adult but would choose to run the car off the bridge to avoid striking an adult carrying a young child. It’s unsettling to think about the many permutations of this problem. While we may be unwilling to hit the family, what if your own children were in the car? The thought experiments on human willingness to trade lives is a long standing one in ethics call the Trolley Problem. While it’s primary historical use-case was for creeping out freshman philosophy students, these though experiments take on a new urgency. While humans are uneasy putting to paper a ranking of how human lives should be value, for a computer to make similar snap judgements that most of us would, we’re going to have to do so, preferably, in a way that’s transparent to everyone.
While AVs have plenty of upside, we can’t allow ourselves to be blinded by the promise and ignore the associated perils. Autonomous vehicles must not, as journalist Alexander Kabakov said over the fall of the Soviet Union, “suffer a victory”. It’s early enough that we can choose to exacerbate the benefits, while mitigating the drawbacks. The good news is, each downside is another opportunity for us to act on this future and build the best possible outcome. Like any new technology, AVs are not inherently good or bad, but an opportunity to change the world. Let’s work to bend that change toward justice for the betterment of more humans.
Denouement
There are reasons to be excited at the possibilities, but heed the reality check of Sertac Karaman, autonomy researcher at MIT: “We may wake up 50 years in the future to find a world where Level 5 AVs never happened.” That said, there are plenty of less fully autonomous transportation options that can still majorly disrupt our economy, which is what makes many of us so confident that AVs will be a major disruptor over the next few decades, from autonomous farming and mining, so distribution centers and highway travel.
We know there are people who like driving. There were people who liked to ride horses. But beyond all of the annual deaths, requiring ownership of a rapidly depreciating asset to take part in the economy is a terrible design. The good news is that Millennials and Gen Z care more about convenience than ownership, so the mere existence of ride sharing is starting to remove car ownership as a requirement for middle class membership in America.
With the changes happening in autonomy and ride share fleets, we’re finally able to step back and ask ourselves if we ever really wanted cars at all, or just the freedom of movement. Safe and available transportation is a cornerstone of liberty in the 21st century and beyond. Our descendants will wonder how we were able to survive this bronze age of human powered transportation, and what took us so long to remove our simian brains from the loop. If we could conquer the world on the backs of a handful of well-trained quadrupeds, we can certainly tame our own machines.
References
No One At the Wheel: Driverless Cars and the Road of the Future — Samuel I. Schwartz
Driverless: Intelligent Cars and the Road Ahead — Hod Lipson, Melba Kurman
