Self-driving Cars Will Drive Better Than Humans, and Sooner Than You Think

This is the first in a series of articles about self-driving cars. The views expressed in this article are my own, and do not represent the views of Travelers Insurance Company.

In 2004 in the middle of the Mojave Desert, an amazing thing happened: fifteen vehicles built by students, universities, and others set off to complete a difficult 150 mile course across the California desert. There were no drivers and no remote controls.  Each vehicle had to complete the course on its own – and not one vehicle managed to complete the course! There were breakdowns, turnarounds, and one vehicle even managed to flip itself over.

(You thought you were having a bad day…)

Thus began the journey toward fully self-driving cars (SDCs). This event, called the Darpa Grand Challenge, was repeated twice more: in 2005, with 6 vehicles making it to the finish line; and again in 2007, this time on a city course requiring recognition of road signs and traffic lights. Soon after, Google and others began serious development and testing of SDCs, and the race was on!

The race is on! Like this, but much, much nerdier.


What is happening today?

Today there are over 35 companies working on this technology, and the current state may be further along than you think. Waymo (Google’s self-driving spin-off company) has been offering self-driving rides to families in the Phoenix area since 2017, and are removing the driver for some trips in 2018. Cruise (purchased by GM in 2017) has been self-driving around the challenging San Francisco streets since 2016. A recent report from the California DMV showed that these two companies experienced disengagements (where a human driver had to take over) at a rate of only 1 per 5,596 miles and 1,254 miles, respectively. While not as far along as Waymo and GM/Cruise, there are also many other serious companies testing vehicles on streets across the United States, including the likes of Ford, Apple, Intel, and Nvidia.


How good are these vehicles today?

I do not work for any of these companies, so my information comes from articles and stories on the internet. My best guess is these vehicles are not horrible, but also not great. Based solely on my own opinion, the better SDCs today operate at about 20% of the skill of an excellent driver. Let’s say this is as good as a poor student driver – and it took 13 years to get here (starting from basically 0% in 2004 at the first Darpa Grand Challenge, up through 2017.) At this pace, it would take another 13 years for these vehicles to be driving at around 50% of the capability of an excellent human driver, with fully capable SDCs still a long way off.


(This chart shows slow, linear growth. Truly capable SDCs are somewhere way over to the right…)


Moore’s Law has something different to say

While the projection above is reasonable, I propose that advancement in SDC capability will progress much, much faster. You’ve heard of Moore’s law, where computer processing speed doubles every two years. This is because progress of information technologies tend to follow predicable and exponential advancement. Also called the Law of Accelerated Returns, this doubling is the reason that we can carry around a 1980’s level supercomputer as a smart phone in our pockets. I like to characterize the affects from this law in three ways: 1) speed/capability increases exponentially, 2) cost decreases exponentially, and 3) size decreases exponentially. Technologies subject to this law are deceptively slow in progress, right up until the point where they seem to suddenly leap forward and advance quickly.


Are automobiles information technology?

Consider the following: new cars today have between 25 and 50 CPUs; they use drive-by-wire technology (which actually means wireless control) to accelerate, steer, and stop the vehicle; they apply sonar, radar, cameras, and lasers for things like adaptive cruise control; they can self-park and self-stop; and very soon they will self-drive. Based on this, cars can very much be characterized as computers on wheels. This means that automobiles (and especially SDCs) would be subject to the Law of Accelerated returns.


(This chart shows exponential growth. Extremely capable SDCs arriving in 2022!)

You might want to accelerate your timeline

Applying this law, I believe there will be a rapid acceleration in the progress SDC capability over the next few years. In fact, according the chart above, SDCs will drive as good as any human driving within 5 years (by the end of 2022). This is much faster than the linear projection of 50% capability in 2030. Beyond 2022, SDCs will continue to advance, driving better than any human, in any situation, in any road condition. Couple this with a corresponding rapid decrease in the cost and size of self-driving technology, self-driving capability would be part of every new car built from this point forward.

If the growth in capability of SDCs does end up following the Law of Accelerated Returns, this technology will take over the automobile industry much faster than anticipated.

About Matt Perkins

I've been actuating for quite a while now... My experience includes #pricing, #reserving, personal and commercial lines, rate implementation, #SAS, #excel, rating systems, leading projects, managing teams, championing others, ad-hoc analyses, presenting to large-ish groups, and lots of other things useful to insurance companies. I also follow autonomous vehicle technology as a hobby and sometimes work effort. I am currently a Fellow of the Casualty Actuarial Society (FCAS). Anything I write here expresses my own views and does not represent the vies of Travelers Insurance Company.

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.