Despite the headlines you’ve probably seen, car crashes still aren’t big news. In the past nine months, there have been four crashes involving autonomous vehicles (AV) since September 2014, according to California’s DMV. Three of the crashes involved Google’s self-driving Lexus SUVs and one was an Audi powered by Delphi. None were serious, or the fault of the technology, and were the fault of inattentive humans – the exact problem that autonomous cars promise to solve.
Last September it became a legal requirement of the Californian DMV for AV manufacturers to report all AV accidents regardless of their severity, even if the car was not in its autonomous mode. With the transition between human and autonomous driving not well defined, and to remove any possibility of a last second switch tricking the system, this last caveat is stringent but necessary.
However, it does mean that 50% of these reported crashes, one of the three Google crashes and the one Delphi crash, were during normal human driving – making them utterly unremarkable. Delphi claims its Audi was broadsided whilst the human driver was waiting to turn left – and so you shouldn’t begin writing off these early trials based on headlines alone.
Google claims its accidents during autonomous driving were the fault of the other, human, party. Google has offered no specifics on the 3 new incidents, or any of the 11 total minor AV crashes it has had since testing began 6 years ago – only stating they were “a handful of minor fender-benders, light damage, no injuries, so far caused by human error and inattention.”
It’s curious that over the last 6 years, 3 of the 11 Google accidents were recorded in the last 9 months, since reporting them became mandatory. However, AV testing by nature involves a lot of mileage over short periods of time; and with testing ramping up, one would expect that the rate of incidents reported to rise – and continue to do so as we move forward.
That said, the norm in the automotive industry is to consider the accident rate per mile, and Google has been quick to dispel early reports that its AV has an increased minor accident rate compared to human driving – predicting sensationalist headlines swaying the public’s, already divided, view on AVs
The national rate for “property-damage-only crashes” is 0.3 per 100,000 miles for human drivers. It has been widely pointed out that, considering Google’s 1.7 million miles of autonomous driving and 11 AV crashes, its figure sits at just above 0.6 per 100,000 miles – twice the non-autonomous average.
Google reminded its critics that as many as 5 million minor accidents go unreported each year – a figure it (conveniently) doesn’t provide any attribution for. Taking this value into account (conveniently) doubles the total reported incidents, and hence doubles the national rate to 0.6 for humans, matching that of Google’s – not that conclusions drawn from a sample size of 11 bear much statistical significance in the first place.
With time, the number of reported AV incidents will rise, and greater analysis into this accident rate will be possible. As there’s always two parties involved in a crash, with blame typically belonging to one or the other, it will be interesting to see where the blame tends to lie for AV incidents. The recorded video and sensor data will also prove particularly amusing, the first time it is used to challenge a lawsuit too.
The technology punters are, of course, adamant that AVs drive safer than humans, and that the majority of fault in future accidents involving AVs will lie with human drivers – and we tend to believe them. The depth and level of attention the AV systems pay to the road and its occupants is much greater than any human -the main limitation currently being the processing of that information, which is only set to improve thanks to the wonders of Moore’s Law.
Early comments state that Google’s AV takes an irritatingly long time to react to a green light, not due to any technical-latency, but for a “better safe than sorry” approach – which makes sense from a damage limitation point of view, but may annoy those stuck behind AVs at the traffic lights. Similarly, this approach explains Google’s recent patent filing for a system that detects livestock that are obstructing the road – and means that the cars will soon be able to tell the difference between a temporary or permanent static obstruction.
Potentially, the over-cautious autonomous driving style will frustrate nearby vehicles which in turn could place AVs at a higher risk of being affected by reckless driving. Unpredictable driving of nearby vehicles is one of the toughest challenges for AVs to deal with, but again, these mobile video platforms should have a wealth of video evidence to hand over to the authorities the first time someone sideswipes an AV in anger.
It’s also unclear how human operated and autonomous vehicles will interact in dynamic situations where right of way has historically been determined by social cues and stern looks – we’re just not sure how Google’s cute AV will assert its dominance. Although perhaps human drivers, unable to read an AV’s intentions, will be overly cautious in these scenarios. Four-way stop signs and roundabouts are a little harder to remotely navigate than light-controlled intersections.
The early years where human and autonomous drivers have to learn how to interact on shared roads will be tricky, no doubt. To a certain extent, we won’t know the subtleties of the interplay between a road-rage driver and an AV until it plays out in full, but for now baby steps need to be taken, and the public will need to keep an open, and forward thinking, mind.
AVs especially need to develop their understanding of drivers and road systems globally, and not just in California or Nevada. As the two sides understand each other better, the work required to coexist on the road should reduce – and then AVs could begin to dial down the hesitance that Google’s trial project is being accused of. It takes society a while to adapt to technological changes, and cars with no drivers is a fairly major transformation.