Future Tense

Should Self-Driving Cars Obey Speed Limits?

The rules of the road will need an update for robocars.

A Google self-driving car is displayed at the Google headquarters on September 25, 2012 in Mountain View, California.
A self-driving car at the Google headquarters on Sept. 25, 2012 in Mountain View, Calif.

Photo by Justin Sullivan/Getty Images

This article originally appeared on Brad Templeton’s site Robocars.

In many countries, including the United States, the speed limit is a rather nebulous thing. It’s posted, but on many roads hardly anybody obeys it. Almost every driver speeds regularly, and anybody going at or below the limit on a clear road outside the right lane is typically an obstruction to traffic—they will find themselves being tailgated or passed at high speed on the left and right.

Half of Germany’s Autobahns have no specific speed limit, but they have a better safety record. For a brief time, Montana’s highways also had no set limit. In France they take a different approach. The Autoroute limit is 130 kph (81 mph), and almost nobody exceeds it; in fact the vast majority go under it. Reportedly, this is because the police are serious about it and will ticket you for any excess.

In the United States, it’s not that way. A ticket for going 1 mph over the limit is an extremely rare thing. It usually signals a cop with another agenda or a special day of zero-tolerance enforcement. In fact, many drivers feel safe from tickets up to about 9 mph over the limit. Tickets happen there, but the major penalties require going faster, and most police like to go after that one weaving, racing guy who thinks the limit does not apply to him.

The limit is a number, but it is not especially magic. It’s not like one is safe at 65 mph and reckless at 66 mph, even though that’s how the law is written. Rather, the risk from accidents increases gradually with speed. The risk of having an accident is harder to measure, but the severity of an accident is related to the square of the speed of impact.

There is a speed at which we may judge the accident risk is above acceptable limits. This speed is not a single number. It varies from driver to driver and from car to car. It varies from hour to hour, from weather condition to weather condition, and from road to road. As the Autobahn’s lower accident rate shows, some drivers are safer at very high speeds on well-designed roads than other drivers are at 50 mph on lesser roads.

And while the Germans are content to do it, the United States is not prepared to officially let drivers decide what the right speed for acceptable safety is. Rather it is done unofficially and irregularly.

So how does a self-driving enter this world of few hard-and-fast rules? There are two common schools of thought:

  1. As with its ancestor, the cruise control, the operator of a robocar can set the car to operate at any speed within its general limits, regardless of the road speed limit. The moral and safety decisions rest with this person.
  2. The vehicle must be programmed to not break the speed limit, nor allow its operator to do so. It must be aware of all limits and obey them.

I believe the first choice is both better and more likely. It’s more likely because the public has a strong love for having control of their cars, even if it is automated. Attempts to put in speed limiters by law have all been rejected, and cars are routinely sold able to go much faster than any allowed speed limit. Some cars feature speed limiters due to European New Car Assessment Program rules and high European ticket costs, and some cars with very large engines have speed limiters that block them from doing truly crazy speeds, though these can be disabled for racetrack use. (Some people also seek speed limiters for children or parking valets, but don’t want them for themselves unless they are one ticket away from losing their license.) In Canada, a truck driver successfully sued to have a court declare a speed limiter put on him was an unconstitutional violation of his basic freedoms.

People generally don’t want their technology to disobey them or enforce the law. It’s similar to the digital rights management battle in music players.

I think it’s more interesting to examine whether limiting vehicles to the speed limit is the right thing to do. At first blush, considering that slower generally means safer, it seems as though such limits would improve safety. This turns out not to be true if, as seems likely, the driver of a robocar can take the wheel and operate it manually—especially if conventional vehicles remain on the road.

A vehicle limited to the speed limit will be going much more slowly than traffic on most U.S. freeways and be forced to drive in the right lane. This is actually a poor lane to drive in, because it means interacting with many cars merging onto the highway. This also means greater variation in speed as cars brake for merges and exits. This is also where the below-limit drivers are found.

This less comfortable ride, plus the longer travel time, will create a great temptation to manually take the wheel on many highways. Here’s where some math comes in. If we presume the robocars have an accident rate that is just 50 percent of the human driver rate, the driver is doubling her risk by taking the wheel. If we presume accident risk is 25 percent higher at 75 mph than at 65 mph, this applies to both the vehicle and the human driver moving at that speed. In fact, it is likely the vehicle would do a better job at the higher speed or it would not be rated for that speed at all. The vehicle will know the exact shape of the road and the capability of its tires, and adjust this as it travels to do a better job.

On many roads all lanes are moving faster than the limit. The limited car would become an obstruction to traffic.

Math of lives and hours

Another interesting piece of math revolves around the reason people speed—they want shorter trips. While a trip in a robocar is not nearly so much wasted time as having to drive yourself, consider that human drivers have a fatal accident every 180 million miles of highway driving (at 75 mph,) and let’s presume robocars get good enough to have one every 360 million miles at 65mph, and one every 300 million miles at 75mph.

Let’s consider time spent driving as 20 percent productive. People have conversations when not alone, do phone calls, and listen to audiobooks and news. Let’s consider time in a robocar as 70 percent productive—you can do many productive things, but it’s not quite as good as being at a desk or at home. These numbers are arbitrary guesses, but you can fill in your own. At 16 hours per day, there are 5,800 productive hours in a year.

If robocars are constrained to 65mph, then 360 million miles of driving takes 5.5 million hours and has one fatality. We have 660 years of productive time saved and 282 years nonproductive.

Humans at 75 mph would use 4.8 million hours and have two fatalities. That’s 82 years productive time and 740 years nonproductive.

A robocar allowed 75 mph also uses 4.8 million hours with 1.2 fatalities. That’s 575 years spent productive and 246 years nonproductive.

In other words, comparing the 65 mph constrained robocar with robocars and people going 75mph:

  1. If nobody takes the wheel manually because they want to go faster, going 65mph saves 0.2 fatalities but costs 36 years of productive time, about half a lifetime.
  2. If everybody takes the wheel to go faster, there are 0.8 extra fatalities and more than 6 full 80-year lifetimes of productive time is spent driving.
  3. If just 20 percent of people take the wheel, we get 1.2 fatalities and 129 extra years of productive time lost.

Clearly the option where people take the wheel is inferior for all. People have many opinions about how they would contrast 0.2 fatalities with 36 years of life wasted. On a pure time basis, 0.2 fatalities probably maps to about 8 years since on average half a lifetime is lost in a fatality. But you may view a death as far more tragic than 80 years of 16-hour days spent wasted, since that is spread out over many people. But there is an argument that even the pure “nobody switches to manual” scenario is not necessarily better, or only marginally better, and the reality is that plenty of people would switch (or simply not get a robocar at all) making the move a clear loss. By the time 20 percent of people are taking the wheel because they are in a hurry, the speeding robocar is clearly the superior result by any metric—based on these assumptions.

Yes, math that puts values on human lives is complex and difficult, though I think that it’s easier to see an equivalence between lost time and lives than it is to see one between dollars and lives.

Nonetheless, I believe the math and other arguments clearly show that robocars should be allowed to move faster than the speed limit so long as they are rated suitably safe in the particular conditions, and the bulk of other traffic is also doing this.

The French system is almost surely better. With the fixed 130 kph (80 mph) limit, there is no issue to resolve. Robocars would be able to easily move the speed of traffic, and people would not disengage just to drive faster than the limit with the strict enforcement.

Efficiency

It should be noted that higher speeds use more fuel, at least at these levels. This is an independent variable, however. People routinely value their time more than the fuel cost and want to go faster, no matter what the means of transportation. If their car won’t go faster and waste energy when they tell it to, they will take the manual controls. The way to solve energy waste is to put the environmental concerns into the price of the fuel.

Minor code violations

Speeding is just one of the code violations almost everybody does. There are many other examples. One that Google reported early on was handling a four-way stop. They found that if they were not a little aggressive in asserting their turn at the stop, others would quickly grab the slot, and the car would sit waiting for a long time. So they programmed the car to try to go on its turn even if somebody else was also trying to steal the slot, though it will eventually yield in a true game of chicken. Fortunately, there is always somebody who will be polite and yield when it was not their turn.

They also programmed the car to do things like nudge a little into the oncoming lane (when safe) to get around things like double-parked trucks.

Another interesting example is the stop sign. The code requires a full stop. So many people instead do a rolling stop it’s called the “California stop” in that state. Usually it results in no ticket unless police have gotten bored and set up a stakeout.

But does a robocar need to always do a full stop? It knows the full geometry of the intersection and where its sensors can see and not see. It can know if it has a full picture of the situation and that indeed there is nobody else within a modest distance of the stop signs, and no pedestrians about to cross the crosswalks in question. In that situation, why not let the car roll through? Unlike a human, the car would only do this when safe.

Another situation is “no left turn” or “no turn on red” intersections. We restrict human drivers because they can’t be trusted. Each “no turn on red” has a reason behind it, and that particular reason can be programmed into the car. For example, there are “no turn on red” and “no left turn” or “left turn on arrow only” locations which are signed this way because too many people will try to take the turn and end up impeding fast traffic. But a robocar that is certain it won’t do this should make the turn when clear and improve the flow of traffic. These signs tend to be in force all the time, not just when there is heavy traffic and they are truly needed. Off peak, there is no reason not to allow these turns when safe and clear.

Testing

Even those who would advocate the cars be speed-limited might think twice when considering the subject of testing. To be safe, robocars must be extensively tested in real world driving conditions, and in the United States, that involves speeding. A car constrained to the speed limit could be tested in only a small subset of real driving situations. While it could be argued that if it never will speed when operating, it does not need to handle such situations, the reality is that speed-limited testing will not adequately test conditions other than driving in the right lane, being passed by most traffic. Keeping strictly to the rules (such as slower traffic keep right), the car will not be able to test extensively situations only found in middle or left lanes, including being passed on the right by speed demons and several other complex flows of traffic. These flows do occur for lower speed cars but less frequently, and still need testing.

This presents another conundrum for the companies doing the testing, which will think twice about the vicarious liability issues of instructing their test drivers to exceed the speed limit, but for safety there can be no other choice.

Going way beyond the law

The current vehicle code is complex but effectively expresses just two principles for moving vehicles:

  1. Be safe.
  2. Do not unfairly impede the flow of traffic.

Most of the regulations are written to help human drivers obey these principles, because they can’t be trusted on their own. Vehicles that can be trusted might be better with an entirely different, much simpler vehicle code. All robocars with the same software will drive similarly, and there will only be a modest number of different software providers in the market.

If a vehicle makes a mistake and violates the code or the principles in a way that the maker will face punishment, the software will be modified and all vehicles will never make that mistake again. This is vastly different from how humans act. If a human makes a mistake, that means it’s more likely that other people will make the same mistake, and the law gets amended to forbid it.

Once robocars can be certified safe, they may well be able to drive even faster than humans can safely drive. While this is not energy efficient, in many ways it might make sense. This could either be done autobahn style, where the left lane has no speed limit and slower cars must pull to the right, or it could involve managed lanes that are limited to high speed robocars at certain times of the day. This would strongly encourage a switch to robocars.

If efficiency is the issue, one could also imagine high-speed lane use being granted only to carpools, vanpools and other non-solo transport. In today’s carpool system, it is hoped that people will do the hard work and suffer the inconvenience of arranging carpools so they can use a lane with less congestion than the main lanes. Only a very few people do this—only a small fraction of the carpools are “induced” carpools; the rest are couples or people already carpooling.

In the robocar world, where your solo car can drop you off to join an ad-hoc carpool and pick you up on the way back, and where a short-range robotaxi can pick you up from where your ad-hoc carpool members part ways, real carpooling becomes far less inconvenient. Given an added incentive like a 100-mph dedicated lane for the bulk of the trip would be quite attractive and save energy to boot.

Even without a dedicated lane, a legal speed limit of 85mph or 90mph in the left lane for robo-carpools could make such travel very attractive, particularly if there is enforcement of a German style rule of leaving the left lane if somebody comes up behind you at a faster speed if safe to do so.

A special driver’s license for the car?

The idea of having a less restrictive vehicle code for the robocar is a challenging one, legally. It will only come after a lot of history. Curiously, the current legal regimes may offer a way to do it faster. The Nevada law, for example, declares that while the vehicle code regularly refers to things the “driver” should do, a self-driving car will be treated as though the “driver” is the person who activated the system. At first that will be the person sitting behind, but not holding the wheel.

The Nevada law also has a special license endorsement. The initial goal of this endorsement was to make sure people who operated the cars were trained in how to operate them. Over time, this becomes less important as the cars will do all the work and in fact we want to allow the disabled to use them.

Initially this endorsement could offer an interesting legal approach. Create a new class of license. This license is valid only when operating an appropriately certified robocar. With this class of license, certain vehicle code restrictions (like speed limits) could be different than they are for people with the general license. This has precedent, though mostly classes of licenses allow you to do things like operate bigger vehicles, or carry passengers for money etc. Why not one that gives you a different speed limit?

This is not a perfect situation. The truth is it’s the car which would prove its ability to safely follow a different code, not the person. But our system of laws is much more tied to people than machines, for good reason.