Your browser is not supported. Please update it.

30 November 2018

Crisis of confidence in autonomous driving safety

The inevitable consequence of being over-hasty to evaluate autonomous cars under more testing conditions is that accidents occur and confidence in their safety declines. So, if autonomous driving is actually safer even with today’s technology that is hard to prove until far more miles have been accumulated under varying conditions. What is beyond doubt though, after a spate of surveys both in Europe and the US, is that public perception of safety has plummeted. That as well as the accidents themselves has caused some major players to decelerate their trials and deployments, with General Motors postponing testing of autonomous vehicles in New York City, after Mayor Bill de Blasio expressed concerns over safety at this stage.

The latest survey, conducted by worldwide travel insurer Allianz Global Assistance, found that just 52% of Americans now would consider trading their conventional vehicles for self-driving ones despite the advantages in comfort and experience, with the rest being deterred by fears for their safety. Other recent surveys have been at least as damning with 73% of American drivers now “too afraid” just to ride in a self-driving vehicle, compared to 63% in late 2017, according to a AAA (American Automobile Association) poll. Only 20% expressed full trust in self-driving.

There is less data from Europe reflecting fewer self-driving trials having been conducted, but feedback on safety seems to have colored views. Japanese car maker Mazda polled 11,008 motorists in the UK, Austria, Belgium, France, Germany, Italy, Netherlands, Poland, Spain, Sweden and Switzerland late in 2017, to assess opinions on various aspect of autonomous driving. Precisely one third of the drivers surveyed welcomed the advent of self-driving cars, reflecting in part conservatism but also lack of conviction that the technology was anywhere near fit for widespread deployment, with safety among the concerns.

Much of the focus over autonomous driving safety has been on California where a lot of the experiments have been conducted but where there is a legal requirement to report all accidents, even those involving the slightest contact with other vehicles, property or people. This means California accounts for a large proportion of reported accidents involving self-driving cars. On this basis Google’s sibling company Waymo has to date been involved in 32 California incidents, representing 41% of the total for its global autonomous trial fleet. Then General Motors’ Cruise Automation self-driving car subsidiary has fared particularly badly in San Francisco, with its autonomous cars accounting for 52 out of 61 accidents involving such vehicles in the city. Among others, robot taxi startup Zoox has had cars involved in five reported accidents, among a handful of others.

Drilling down into these reports two points stand out. Firstly, nearly all these accidents were the fault of the other vehicle, which in turn begs the question of whether the self-driving system could or should have reacted better to the emergency with evasive action. As insurers are well aware in most accidents both parties are partly to blame, for while it might be only one making a serious error the other might have contributed through lack of concentration or poor reaction.

This links to the second point which is that most of the accidents have occurred on urban or suburban roads where there are many more hazards and challenging conditions. Autonomous cars already have a demonstrably excellent record on quiet open highways, but then so do human drivers apart from the risk of falling asleep at the wheel.

However, there is accumulating evidence that at present humans have the edge in congested suburban conditions, especially where vehicles are still driving at some speed and face multiple hazards at various ranges. Human drivers tend to improve up to a point as they gain experience of combining concentration on the road immediately around them with awareness of potential dangers lurking further ahead. They may notice at the edge of their vision a vehicle approaching a junction faster than usual for example so that it might not stop in time.

Self-driving vehicles are as yet incapable of incorporating such feedback from distance, which is why there is work on deeper reinforcement learning algorithms that can take account of more diverse visual aspects within a vehicle’s viewing horizon. Meanwhile though the message for the self-driving industry may be that more work is needed in controlled conditions before pushing ahead with large scale urban tests.

There is irony in the fact is that the one proven safety benefit of vehicle telematics is not directly associated with the technology itself, but just knowing that it is there. This is behind a sharp reduction in number of young drivers injured or killed as a result of having black box monitoring for insurance.

Almost under the radar there has been a boom in telematics policies among young drivers with almost 80% of those aged under 25 having these in the UK now. They have been motivated simply by the high insurance premiums they have to pay if they are under 25, and the substantial savings they can make if they have a black box installed in their car.

The mere presence of this box has yielded a dramatic fall in accidents among this age group, admittedly from a higher base then older drivers. The most comprehensive survey so far conducted in the UK by LexisNexis Risk Solutions shows that the number of 17-19-year-old drivers who have been killed or seriously injured in road traffic accidents has fallen by 35% since 2011, compared to 16% for the driving population as a whole.

This comes despite negative feedback about the telematics technology itself. Nearly seven in 10 drivers do not trust ‘black box’ technology to lower their insurance premiums by demonstrating it is safer, according to research by This partly reflects fear of Big-Brother-like intrusion unto privacy and partly the hubris most evident among male drivers that they are already as safe as they can be.

But it also reflects the primitive state of the technology which is often based just on GPS tracking, in which case it is incapable of deriving accurate data about breaking, accelerating and steering, being confined largely to fairly rough measurement of stopping distances according to speed and the gap to the vehicle in front. The boxes are not smart, and often struggle to differentiate between a bumpy road and erratic driving.

Furthermore, scoring mechanisms do not yet take account of behavior that reduces risk of accidents, such as good anticipation and intelligent evasive action resulting in near misses rather than actual collisions. Some of this will be remedied by more advanced devices that also incorporate OBD (On-board diagnostics) data captured directly from the vehicle.

Even so telematics insurance for the young is a good news story that has not yet received much publicity and there is every reason to expect that these benefits will ripple across the whole driving population. It will be some time before similar conclusive evidence of safety benefits from self-driving comes home.