To the general public, the term “self-driving vehicle,” as popularized by Google in the early 2010’s, conjures images of people riding to work deeply engrossed in books or phone calls, oblivious to the outside world – no different from being chauffeured, picked up by a taxi, or riding on a public bus. However, because that level of automation will certainly come last in a long evolution of incremental improvements in AI technology, we are likely to be driving, and then be driven by, vehicles which are not fully automated, but instead are responsible for more driving tasks than they are commonly required of today. So, then just what do we mean by a “Self-Driving” vehicle? And furthermore, how can we understand what the driving environment will look like when our streets are populated by a mix of driverful and driverless cars?
SAE International has developed a 6-tier system for describing the levels of vehicle automation and has become the de facto standard for contextualizing what we mean when we talk about “self-driving” cars. In fact, in the first example given above, SAE would say those people enjoying their book unhindered by the task of driving were enjoying a ride in a Level 5, Full Automated Vehicle (L5-FAV). At this pinnacle of automation, the driving system alone is responsible for the full-time performance of all driving tasks in all conditions and environments, behaving exactly as a human driver could and would. According to SAE, this includes steering and accelerating, or watching and adjusting to the driving environment. Perhaps most importantly, an L5-FAV itself is the fallback agent; lower levels of automation are defined by the necessity of human intervention in novel or highly unusual situations – the Agent on whom the system falls back in order to complete that task is the human driver themselves. In an L5-FAV, human intervention is not necessary – the system is so advanced that it and it alone is the fallback for all driving decisions.
As such, the benefits typically enumerated by AV advocates tend to be drawn from the L5-FAV level – a reduced number of vehicular accidents, greater mobility for the visually impaired or disabled, increased human productivity on the road while commuting, etc. We can reasonably conclude that such a level of automation is not only preferable, but also possible. Artificial Intelligence has advanced significantly in recent years; it goes so far in some cases as to demonstrate robots which can simulate childhood learning using a trial-and-error algorithm. In 2014 researchers at Google DeepMind developed an AI system that learned how to play Atari 2600 video games, by remembering the results of past experience, analyzing them, and then modifying its own code in order to maximize its video game score. This type of learning has clear implications for driverless cars. If Google DeepMind could learn to win a racing game like the popular Need For Speed 2015, for example, then in theory it could learn to drive a real vehicle. If the futurist Ray Kurzweil (also a Google employee) is correct that by 2029 AI will be as intelligent as the average adult human, then driving a car should easily be within the realm of possibility. However, it is important to remember that as Full Automation will be the final stage of a mature technology, it is virtually certain that in the race to the market place, Automated Vehicles at more rudimentary automation levels will be introduced long before Full Automation is a possibility.
This is where some of our greatest ethical dilemmas in using AVs start coming into play – when we have a mixed system of vehicles of all different levels of automation. Currently, what we lack is a robust theoretical model similar to SAE’s Levels of Automation, which describes the stages of the driving environment’s evolution as automated systems account for ever greater percentages of the overall vehicles on the street. We can guess that a Level 0 would be no automation in the any driving system (primarily, what we have today), and at the opposite there would be – some level n – in which all vehicles were fully automated and the benefits of AV’s were fully realized. What we don’t know is how many useful stages of progress there would be in between those extremes.
To begin addressing this problem, I have modified a graphic originally posted on Vox.com and created for its own article on Driverless vehicles. The graphic is derived from SAE’s standards for levels of automation. Here, I propose a simple reference system for distinguishing types of driving environments based on the mix of driving automation levels on the road in a given area. I propose three broad categories of Driving Environments: Type A0, Type B#, and Type C4/5.
Under this classification system, a Type A0 Driving Environment would designate a clearly-defined corridor, highway system, or geographic area (preferable in lat/long coordinates, or using the US Military Grid Reference System) in which no automation is present in any vehicles. At the exact opposite end of the spectrum, a Type C4/5 Driving Environment would be one in which all vehicles present in the designated area would be either highly or fully automated. But it is the middle, Type B#, where this taxonomy system could prove most useful. A TypeB# driving environment would denote a mix of vehicles at varying levels of automation, following the numbering system used by SAE. Under this system, a Type B01 DE would refer to those roads on which there is a mix of traditional and Driver-Assisted vehicles, while a Type B015 would mean that at least some of the vehicles operating in the area would be fully automated, some would provide simple driver assistance, and still classic, driverful cars would also be mixed in and remain on the road.
The efficacy of this system is based on the assumption that it would be legally prescriptive, and not merely descriptive; we intuitively know that a TypeB# driving environment would behave differently from an environment consisting solely of L5-FAVs, A0’s, or even one that managed to be exclusively of Type B3 – yet despite intuition, our legal system doesn’t know that yet. The Driving Environment classification system I have proposed here is therefore intended as the basis for a common legal framework for designating AV zones. This is so that lawmakers could exercise some zoning control over such a new technology, effectively “sandboxing” it until the technology demonstrated enough common weaknesses and strengths as to allow the timely shaping of norms in the insurance industry, policing, and the court systems.
For one such example of possible application, in many large cities today local governments have designated High Occupancy Vehicle lanes and/or Zip Lanes which are often physically separated by concrete barriers from other lanes of traffic moving in the same direction (think: Washington D.C. or Honolulu). If a nuanced classification system were adopted, city governments could easily designate, say, a TypeC4/5-Only lane during typical commute hours. That common framework could be just as amenable to tourist destinations that wish to designate a TypeC-Free Zone in an historic square or monument; any vehicular accident occurring in such a zone between at Type A0 and a L5-FAV could then be easily decided in favor of the Type A0, since the L5-FAV’s owner was never legally allowed to send the vehicle there in the first place. Hence, by adopting common definitions which adequately recognize the nuances in the varying levels of automation, lawmakers would be empowered to write legislation with greater granularity. By that token, police officers, insurance companies, and lawyers would be able to make clearer arguments when determining fault in an accident which involved an AV.
Finally, the proposed system I have presented here is not intended as a panacea for some of the largest, thorniest, most intractable ethical issues associated with Automated Vehicles. While a common language would certainly benefit in the ways described above, it is clear that still the social, criminal, and legal solutions to living in a TypeB# society would necessarily vary based on locality, specific AV programming, and by the legal precedent set when the first lawsuits involving AVs inevitably appeared in courts and in the public consciousness.
In part II of this blog series, The Case for Automated Vehicles, we will take a closer look at the promises held for Automated Vehicles and a society transformed by them, while suggesting how some of those ethical issues might be addressed with software, sensor technology, and Federal law.
In Part III, The Case Against Automated Vehicles, I play Devil’s Advocate and do my best to tear apart every argument made in Part II. I will argue that it doesn’t matter if you believe Automated Vehicles should be pursued – we as a people will never allow them to develop to the point that they are viable.
In Part IV, The Song of the Open Road, I take a decidedly lyrical approach to the topic of Automated Vehicles, examining what it might mean for our identity as Americans and the soul of our country were we to give in to automation and give up our love of the open road and the great American Roadtrip.