Autonomous vehicle safety myths and facts, 2019 update

It has became a tradition that I write a quick update on the state of self driving car development every year when the California DMV releases their disengagement data [ 2017 post here, 2018 post here]. 2018 was an important year for self driving as we had seen the first fatal accident caused by an autonomous vehicle (the infamous Uber crash in Arizona).

Let me start with a disclaimer: I plot disengagements against human crashes and fatalities not because it is a good comparison, but because this is the only comparison we have. There are many reasons why this is not the best measure and depending on the reason the actual "safety" of AV may be either somewhat better or significantly worse than indicated here. Below are some of my reasons:

  • A disengagement is a situation in which a machine cannot be trusted and the human operator takes over to avoid any danger. The precise definition under California law is:
    a deactivation of the autonomous mode when a failure of the autonomous technology is detected or when the safe operation of the vehicle requires that the autonomous vehicle test driver disengage the autonomous mode and take immediate manual control of the vehicle.” Section 227.46 of Article 3.7 (Autonomous Vehicles) of Title 13, Division 1, Chapter 1, California Code of Regulations.
  • Some companies appear to report every disengagement, even the benign ones (say to bypass a stuck garbage truck) that do not directly relate to safety of the operation but only to convenience. As far as I'm concerned, Waymo and perhaps a few others don't include those in their numbers. Their reported numbers are only directly related to safety of the operation (as defined by the rule above), hence any potential such disengagement if not for the human override may have, with reasonable probability, directly endangered either the AV itself and its passengers or other vehicles or pedestrians. 
  • Looking at crash-only numbers is not a good measure of AV safety since this number measures the safety of combined AV+safety driver (and in some cases even driver + extra engineer). It is rather clear that modern, well maintained vehicles filled with high tech driven by attentive, trained drivers in familiar neighborhoods will have much higher average safety than ordinary car driven by an average Joe in a shady part of town. 
  • Disengagement numbers may be relatively easily tweaked. Some company can for example spend inordinate amount of miles on a single simple road and crank up a bunch of disengagement free miles. E.g. this year Delphi had zero disengagements but they have only driven less that 300 miles. Hence I did not include their number in the plot below. In fact only companies that had driven significant number of miles matter, these are Waymo, Cruise and Zoox and perhaps Apple.
  • Given this data is provided voluntarily by the companies (many companies decide to focus their testing in states where such reporting is not necessary), we should expect that if anything these numbers represent a rather an optimistic state of affairs. It is likely that testing in adverse conditions such as bad weather where disengagements may be a lot more frequent can be easily conducted in other states where such detailed reporting is not required.
  • Since the reported disengagements are directly related to safety of operation (at least of Waymo), it is reasonable to assume that significant portion of these incidents would have resulted with a crash if not for human intervention. It is likely that some of these crashes would have resulted in fatalities. What exactly would those fractions be is not known, but my personal choice is that one in three of the disengagements reported would have resulted in a more or less major crash and one in thirty would result in a fatality. These are just my fractions, you may choose yours depending on how you feel about it. Also note that for companies that report every disengagement whether safety related or not, these fractions should likely be adjusted by a few orders of magnitude.  

So here is 2019 data (which is a report for the year 2018) - click to enlarge.

This data came from

  1. California DMV disengagement reports for years 20182017, 2016 and 2015
  2. Insurance Institute for Highway Safety fatality data.
  3. RAND driving to safety report.
  4. Bureau of Transportation Statistics

all which is easily verifiable. 

Comments

First of all, this field has became pretty crowded, there is now 48 companies that have reported! Most of them however are startups that did not drive many miles and had disengagements every few miles. A handful of big companies reported zero miles driven (they likely tested elsewhere) or some very small number of miles (less than 10k).

The trend shows improvement from last year across the board. However this is still far from the safety necessary for deployment. E.g. Waymo now reports one safety related disengagement every 11k miles. As impressive as this number is, assuming my fraction of 1/3 of these events ending with a crash, that would indicate a crash every 33k miles. E.g. for my family, since 2014 we have driven some 60k miles across 2 cars. If we were to go with these numbers we'd have to have cause almost two crashes over that period, which we thankfully did not. In fact since I started driving many years ago I had not caused a single crash, much like a great majority of human drivers.

Notable is the appearance of Apple. Their data is however hard to plot since they changed the methodology of classifying their events mid year. Before that they reported everything and that number was literally an event per mile. Since July they've driven some 56k miles and reported 28 events they now consider important (3 of which were actually related to collisions with other traffic). Anyway, that would be roughly 0.5 disengagement per 1000 miles which would put them at the same level as Zoox.

Baidu had driven less than 20k miles using only handful of vehicles. When a big company does their testing at such small scale I consider these efforts not very serious. An R&D project of small/medium size but not really anything aimed at commercialization anytime soon.

Uber had their testing going until the accident in March when they've effectively stopped all testing. Their numbers until then were not particularly impressive with several disengagements reported per each mile. Notably though, Uber did not do the filtering that others are doing so it is not a fair comparison really.

Last but not least is Tesla. They had reported zero miles driven in testing conditions for a second year in a row. Again much like last year they argue that they drive millions of miles in shadow mode on autopilot. I'm not impressed by this. They may indeed generate some data with their deployed vehicles but they need that data to be labeled to be of any use. Tesla cars do not have lidars and rely only on several cameras, a few ultrasonic sensors and a radar for their perception. As I've mentioned in a post several months ago, Tesla took down their Full Self Driving option from website without much explanation. And while they are collecting some unique data on SAE Level 2 semi-autonomy (as nonsensical as it sounds), I don't think that data is particularly useful for Level 5 autonomy which they promised and already started selling two years ago.

And while we are on the subject of Tesla, it turned out that the famous study in which NHTSA concluded that Autopilot improves safety was fatally flawed. A more thorough analysis of the data which had been finally released after few year barrage of FOIA requests showed that cars equipped with Autopilot get significantly more airbag deployments than cars without, from which it can be firmly concluded that autopilot actually decreases safety. I think this case is a great scandal, particularly embarrassing to the government agency undermining trust in any further governance over this new technology. You can read more details about this scandal e.g. here.

If you found an error, highlight it and press Shift + Enter or click here to inform us.

Comments

comments