Every day in the United States, commuters, parents, children, and workers enter their vehicles and travel an astonishing 8,750,000,000 miles per day (for those counting, that’s 8.75 billion miles per day). Most of those trips are routine — going to the grocery store, a doctor’s appointment, or to the office. But for roughly 40,000 people per year, that will be the last trip they ever take.
Road fatalities in the US reversed their gradual decade over decade decline starting in the early 2010s (texting and driving anyone?) and have settled around that 40,000 number for the past several years. That’s about 110 people every day that lose their lives on the roads. I believe that in the 21st century, we can make road fatalities as rare as getting struck by lightning (300 people per year), but doing so will require a massive amount of coordination, safety testing, and societal adjustment. The dream of autonomous, perfectly safe vehicles that do not crash is attainable, and a worthy goal we should strive for.
A Brief History of Autonomous Driving
The first inklings of desire for a driverless future arose in 1925. An electrical engineer named Francis P. Houdina rigged a vehicle with motors and a radio antenna that allowed him to control the speed and direction of the car remotely. General Motors developed Futurama for the 1939 World’s Fair. This exhibit and ride correctly predicted a vast, interconnected highway system that came to fruition through Eisenhower’s commitment to federal highway construction. Unfortunately, radio-controlled automatic highways did not. George Jetson hopped in his flying car, punched in his destination, and was autonomously whisked away to work.
If GM had access to AI image generation tools, maybe this is what “Futurama” may have looked like
It wasn’t until about the 1980s when the dream of a self-driving car inched its way forward on the spectrum of possibility. Teams from Carnegie Mellon and Mercedes Benz created vehicles that could self-drive under certain conditions. In 1995, another car built at Carnegie Mellon completed 98% of a cross-country road trip without human intervention. In 2004, DARPA created a Grand Challenge competition that invited participants to build autonomous vehicles to navigate a 150-mile course. The best performing vehicle completed 7.32 miles in the inaugural edition of this race. The next year, a team from Stanford University unleashed Stanley (pictured below) on the course, and claimed victory, finishing in 6 hours and 54 minutes.
Stanley was a diesel Volkswagen Touareg equipped with rooftop LIDAR units, an electric motor to control the steering wheel, and a hydraulic piston to shift gears.
Excited by the promise of self-driving, and inspired by the successes in the DARPA grand challenges, Google launched their own self-driving car project in 2009. Tesla introduced “Autopilot” in 2014 which enabled lane-centering and speed control without driver intervention. As competition in the sector ramped up, the first of fatality involving a self-driving car occurred. In 2018, a pedestrian named Elaine Herzberg was struck and killed when an self-driving Uber failed to detect her walking a bicycle across a highway. A five year legal battle ensued, with Uber ultimately being cleared of criminal wrongdoing, and the supervising human driver pleading guilty to endangerment. The case made national headlines due to the uniqueness of the event and the ethical concerns regarding it. More on that later.
Fast forward to today, and the state of autonomous driving has continued to advance. Driverless Waymos inhabit the streets of Los Angeles, Phoenix, San Francisco, Phoenix, and Austin. Just last week, Tesla rolled out it’s long-hyped (In 2019, Elon Musk predicted a million robotaxis on the road by 2020) robotaxi service in Austin as well. The autonomous driving future isn’t here, but the seeds are planted, and cultivation is ongoing.
How an autonomous future comes to pass
Now that we’ve got a brief history out of the way, it’s important to explore how the technology is categorized and how it works today, what the differing approaches amongst competitors are, and how these approaches and strategies might evolve in the future to deliver on the lofty goal of fully autonomous driving.
In 2014, SAE International, an automotive standardization body, published an initial classification system that aimed to codify a spectrum of autonomous vehicle capability. The latest version, updated in 2021, is pictured below.
SAE J3016 Levels of Driving Automation
It’s useful to have this reference available when thinking about the progress made on self-driving so far, and where it is headed in the future. Many new cars today come with features that would classify as SAE Level 2, such as lane assist, adaptive cruise control, and brake assist. So you might even have experience with autonomous driving today – you just didn’t know it was classified as that. Currently, companies like Waymo and Tesla are focused on developing Level 4 autonomous driving. Some characteristics of these self-driving vehicles are operation within a specific, pre-defined geofenced area and the lack of a human driver behind the wheel.
When it comes to the technology stack that companies are using to pursue autonomous driving, there are two basic approaches – Tesla (Camera Only + AI) vs. Waymo (3D mapping + Camera + Lidar + Radar + AI), outlined in the graphic below from Bloomberg.
As you can see, and probably surmise, Tesla’s approach is far more scaleable and cost-effective. The sticker price for a Waymo vehicle is around $180,000, the high cost of Lidar and Radar units contributing significantly to that amount. Additionally, Waymo relies on highly detailed 3D mapping of the geofenced area in which it operates. So Tesla has an advantage when it comes to cost and scalability, but will losing the additional sensor information gained from Lidar and Radar, and operating without a 3D map for reference, reduce the overall safety of autonomous Tesla Robotaxis? I think it’s definitely too soon to tell definitively and to what degree, and I also think there’s more to the safety story than just statistics.
Safety Statistics, Failure Modes, and Human Factors
Beyond the enormous total addressable market for taking over the role of the human driver, and the astounding economic value that could potentially be captured, there is one outcome of a driverless future that is unassailably “good” — reducing road fatalities to zero. Autonomous driving, when rolled it in a responsible way and operating under conditions that are appropriately constrained to the technological ability of the system, is already safer than human driving. Waymo recently released a detailed report that provides a comprehensive overview of the safety benefits achieved by their automated fleets.
Waymos are already safer than human drivers when comparing accident rates over mileage travelled
These results are very promising. Who wouldn’t want to live in a world where we could reduce crashes by 90%? Impressive as they are, these statistics come from an extremely small “sample size” when compared to the total vehicle miles travelled each day, and rely on an expensive, gold standard technology stack that incorporates data from multimodal sensor arrays and detailed 3D mapping of the areas in which they operate. It will be very interesting to see the safety reports around Tesla’s Robotaxi offering, as that system relies solely on camera input and AI systems to pilot the cars.
Statistics, however, are not the only piece of the puzzle when we think about how these autonomous vehicles are going to be able to be integrated into our lives. An interesting phenomenon that I’ve observed that’s going to have an outsized impact on the general public’s appetite for accepting self-driving cars is the fact that the “failure modes” for these autonomous vehicles are sometimes non-sensical. Because these systems operate completely differently from a human driver, sometimes when they make a mistake, they make a mistake that a human absolutely would not make. I’ve collected a few examples below.
A Waymo speeds through a flooded sinkhole, completely ignoring a public works crew that was attempting to block off the scene and redirect traffic
A Tesla Robotaxi fails to stop when a UPS truck begins to back up, prompting the safety monitor to stop the car.
Another Tesla Robotaxi slams on the breaks twice when it notices police cars on the side of the road
These are three examples of behavior that depart completely from the way an attentive human driver would handle these situations. Even novice drivers would know to stop or adjust their course when confronted with a public works crew guarding a flooded sinkhole, apply the brakes when a vehicle begins slowing down and then backing up in front of them, and realize that stationary police cars on the side of the road not impeding traffic is not cause for slamming on your brakes.
There is already a cottage industry popping up of collecting and sharing these autonomous driving fails. The Verge compiled a list of these events and even the relatively pro-autonomy Self Driving Cars subreddit is keeping track. In an effort to stay as neutral as possible, I won’t condemn this behavior — I actually think it’s really important to collect data on these failure modes and to spread awareness of them to prevent the technology from rolling out before it’s ready for primetime. Despite this, I also think it’s going to present a very difficult challenge to appropriately frame these failures against all of the safe, successful miles that these cars drive, as evidenced by Waymos safety report. In keeping with the old adage of “If it bleeds, it leads”, depictions of these failures are much more likely to be “newsworthy” than a boring summary of safety statistics. Add to this the fact that there is a lot of social media clout to be gathered by dunking on AI, we’re far likelier to see and be moved by videos of these failures than overall safety metrics.
A final point that’s important to explore, one that’s slightly related to the proliferation of and appetite for these failure videos, is the fact that the way humans perceive and understand information is going to significantly impact the acceptance of self-driving. People are not naturally good at effectively and dispassionately assessing risk, disconnecting their personal feelings and beliefs from the reality of the situation. A prime example is the fear of flying. You are far, far, far more likely to die in a car accident than you are in a plane crash, but have you ever heard of someone who is afraid of riding in a car? Probably not.
There are a lot of reasons the that fear of flying exists. Every commercial aviation accident drives worldwide headlines, plane crashes are more likely to be fatal than car accidents, and when you’re flying commercially, you have no control over the situation. These facts drive emotions and perceptions about the safety of driving vs. flying, and no matter how many statistics you cite like deaths-per-passenger-miles, people are still going to be afraid of flying. I don’t think self-driving advocates are going to effectively convince people about the benefits of this technology by statistics-spamming.
The Good Future and How to Get There
I’ve spent much of this post discussing some of the pitfalls and challenges that face self-driving cars today. I think it’s really important to be intellectually honest, and handwaving the current state of the technology, warts and all, is intellectually dishonest. I want to conclude this post though by talking about how I think we get to the best version of the future and what that good version of the future might mean.
First, I think it’s absolutely imperative that the federal government embarks on building a regulatory framework that’s based on the SAE levels of autonomous driving. I don’t think it should preempt the local experiments that Waymo and Tesla are doing in various cities around the country, but I think it’s going to be important to lay the ground work for federal regulations around self-driving cars. If I could wave a magic wand, I’d want a huge portion of the Nevada desert to act as a self-driving proving ground, consistently incorporating new edge cases and learnings from self-driving fails to iteratively improve the technology.
Additionally, speaking from a magic wand standpoint, I’d want to start collecting video data from the millions of cars on the road today. There are billions of miles of data produced every day that would be hugely beneficial in training vision models, identifying edge cases and behavioral preferences in the vast problem space of driving. This would possibly be a privacy nightmare, and I don’t know how you could implement it effectively or ethically. Perhaps something like an insurance company offering reduced rates if the vehicle collects this data. Again, keep the magic wand, rather than the panopticon, in mind.
Finally, I’d love if everyone would just become a bit more neutral about this topic. That’s kind of the point of Clearly Intelligent, and I know it’s going to be a hard fought battle. But if autonomous driving companies could spend less time talking about eliminating all human drivers in the next 10 years, and the public could break the status quo bias of accepting 40,000 road fatalities in a year in the name of ‘keeping humans in charge” of driving, we might actually be able to chart a path forward.
The good future I envision not only involves rare-as-lightning-strike road fatalities, but also redesigned cities, with more plentiful housing and denser, richer communities. A future where car ownership isn’t a necessity to simply exist in many parts of the country. A future where clean, safe, reliable transportation options change the way we move around.
There’s much more to say on this, and it’s a topic I’ll be coming back to regularly. Advanced technology helps us reframe what’s possible and helps solve major problems that exist in the world. Autonomous driving is a perfect example of this, and even if the road we’ll travel is bumpy, there’s a good future we can achieve, as long as we are intentional and measured in pursuit of it.
Leave a Reply