March 19, 2018

Are Uber’s autonomous vehicles safe?

This morning I learned of the tragic death of a pedestrian in Tempe, Arizona last night, who was struck by an automated vehicle (AV) being tested by Uber. The obvious question this raises is whether AVs in general and Uber’s AVs in particular are more or less safe than conventional vehicles. The problem is that there is really not enough data yet to know for sure. As a RAND report from last year showed, AVs would have to log billions of miles of operation and/or testing before we could draw decisive conclusions about their safety. In mid-December, Uber reported that they had accumulated 2 million miles of driving, and were logging about 84,000 miles per week. Extrapolating this to mid-March, we can estimate that they are only around the 3 million mile mark now.

Nevertheless, one thing we can look at right now is the probability of a fatal crash occurring at this point in Uber’s testing, if its AVs were equally as safe as manually driven vehicles. We can assume that crashes follow a Poisson process – that is, they occur at random points in time, independently of when other crashes occur. In a Poisson process, we expect the time from the start of some observation period (e.g. the start of testing) until the occurrence of a first event (e.g. the first fatal crash) to follow a negative exponential distribution. The cumulative distribution function for the negative exponential distribution tells us that for a rate of 10.7 fatal crashes per billion miles (34,439 fatal crashes in 3,220,677 million miles, per IIHS), there would be only a 3.2% chance of the first fatal crash occurring in the first 3 million miles.

Another simple way to look at this is that in the U.S., manually driven vehicles have one fatal crash every 93 million miles, on average. Uber’s AVs made it to just 3 million miles before their first fatal crash.

Update:

Some people have asked if it wouldn’t make more sense to consider autonomous miles by all manufacturers, not just Uber. On the one hand, this makes sense because we would be having this exact same conversation if it had been a Waymo van that had been involved in a fatal crash. On the other hand, I also think it makes sense to focus on individual fleets, since there are presumably significant differences in the underlying software.

In any case, we know that Waymo had accumulated 4 million miles by the end of 2017, including 2 million in 2017 alone. Extrapolating this through the first 2.5 months of 2018, we can estimate 4.4 million miles for Waymo through mid-March 2018. All other companies combined accounted for less than half the miles Waymo accumulated in CA in 2017. We might therefore assume that all companies other than Uber and Waymo account for a cumulative 2 million miles. Under these assumptions we would estimate the total autonomous miles to date to be 3 + 4.4 + 2 = 9.4 million miles. Having your first fatal crash in 9.4 million miles still doesn’t look very good, considering that human drivers average 93 million miles between fatal crashes. And applying the negative exponential distribution, we find that the probability of the first fatal crash occurring within 9.4 million miles would only be 9.6% if AVs had the same fatal crash rate as human drivers.

Another issue: Engadget reported at the end of last year that

Waymo drove 352,545 of those miles in The Golden State from December 2016 to November 2017. Within that period, the company reported a total of 63 disengagements (instances wherein the human test driver had to step in), which means its vehicles drove an average of 5,595 miles for every disengagement.

The elephant in the room here is What would have happened in those cases if the human test driver had not stepped in?

Assume for the moment that 90% of those disengagements would have turned out more or less ok even if the safety driver had not intervened, and only 10% would have led to police-reportable crashes. That would still be one crash every 56,000 miles — about 9 times higher than the average crash rate for the US in 2015. (3.11 trillion VMT / 6.30 million police-reported crashes –> one crash every 490,000 miles).

Again, it’s too soon to draw firm conclusions, but this does tell us that a crash occurring this soon would be pretty unlikely if AVs were as safe as manually driven vehicles.