Teach Humans to Drive before Robots

Sometimes it seems like we have already given up on human driving. We hear the news about annual highway fatalities on the rise (in the U.S. and globally) and the billions of dollars being invested in autonomous driving and just assume that the car companies or the universities or the folks in Silicon Valley or maybe Baidu or Elon Musk, someone or some organization is going to figure out how to replace humans behind the wheel...and fast.

Spoiler alert! We have a couple decades yet, at least, of human driving so it’s time to take up the challenge of helping, no making, humans drive better. Research tells us it can be done.

This is especially important because we are depending on humans to teach the robots how to drive. The reality is that we really don’t know how good or bad humans really are at driving or even whether they are getting better or worse.

The only reliable reference is the failure rate – i.e. the number of annual fatalities and injuries. But these rates vary by region and over time and are influenced by the implementation of new technologies in and around cars along with laws, licensing procedures and enforcement.

So what is the use of having humans teach robots to drive if humans are actually terrible drivers? Our collective gut feeling is that humans are pretty good drivers. The research indicates that humans can be taught and trained to do much better – and without a lot of effort. But policy makers have to embrace this responsibility as fervently as they are seeking to regulate autonomous driving.

In other words, legislation is being written to manage autonomous driving while the modification, if not the perfection, of human driving is being wholly neglected. Something has to change.

How can legislators and regulators help? Is there really any point to modifying licensing requirements or driver training? Yes, there most certainly is.

A summary of such research published a year ago in Current Directions in Psychological Science (CDPS)– “Hazard Perception in Driving” – describes the process via which drivers can be taught to better anticipate and avoid driving hazards. The report does not sugarcoat the challenge in question noting:

“The problem is that it is also a skill that appears to take decades of driving experience to acquire. This raises the question of whether it is possible and practical to accelerate this learning process via assessment and training in order to improve traffic safety.”

These words reminded me of the estimated billions of miles that autonomous vehicles would have to drive before statisticians could say with a 95% confidence level that the robots can be counted on to drive at least as well as humans (measured by current rates of fatalities and injuries from crashes). In its own study of studies, the Rand Corporation concluded in 2016 that autonomous vehicles would have to drive 11B miles to demonstrate with a 95% confidence level that their failure rate was 20% better than humans.

SOURCE: Rand - For more information on this publication, visit www.rand.org/t/RR1478 - "Driving to Safety"

In the absence of billions of driven and recorded miles, Rand concluded that simulation or other forms of modeling (from the likes of Cognata or Metamoto) will have to be applied before any claims for the superiority of robot driving can be verified. So we humans need to get to work.

The CDPS report highlighted the methods used in the United Kingdom, where a hazard-perception test is included in the licensing process producing an estimated 11.3% reduction in non-lowspeed crash rates. The report goes on to describe the different means of hazard perception testing including:

• In one approach, participants deliver a running commentary about what they are attending to while watching a video clip of traffic and/or listening to an expert driver give a commentary;
• Another approach is to stop a video clip just before a hazard and ask participants what they think will happen next, before presenting them with an expert’s commentary on what might happen next, followed by showing them what happened;
• Finally a combination of the two approaches resulted in a bigger training effect.

The key takeaway is that these interactive approaches have produced results that have stood up over time with driver participants demonstrating enduring retention of the prophylactic effect. Less effective have been more passive, classroom-based exposures and descriptions of hazardous driving conditions. Moreover, the interactive method has been shown to be effective while requiring no more than 35 minutes of participation.

Are there contradictory findings within the voluminous research into driver hazard detection? Yes. Is there more research to be done to refine these methods? Yes. Do distraction, fatigue, alcohol consumption and speed remain overriding fatality and injury-producing challenges? Yes.

The point is that we have more control over and better understanding of the human behind the wheel while we work on perfecting the robot, waiting to take over. Let’s not give up on the humans because, in the end, it is the humans that will have to teach and or program the robots how to drive.

A final note: For decades, automakers blamed humans for highway fatalities. It wasn’t until the 1970’s and ‘80’s and the onset of seatbelts and airbags (long-resisted by the industry) that carmakers recognized they had a responsibility to make cars safer.

Since that time further passive measures have been added – stability control, anti-lock brakes, automatic emergency braking – contributing to a long-standing (though recently reversed) decline in crash fatalities and injuries. The point, though, is that technology and safe driving are inextricably linked and the measurable driving skills of humans, based on those changing injury and fatality rates, is a moving target.

The proliferation of features such as adaptive cruise control, blindspot detection, lane-keeping assist, cross traffic alert, traffic jam assist and, within a few years, vehicle to vehicle and vehicle to infrastructure communications will themselves require a reconsideration of driver training and licensing procedures. The better than 10% reduction in crashes seen in the U.K. would potentially translate to the saving of more than 3,000 lives annually in the U.S. It is worth noting that the fatality rate per 100M miles driven in the U.K. is one of the lowest in the world.

Roger C. Lanctot is Director, Automotive Connected Mobility in the Global Automotive Practice at Strategy Analytics. More details about Strategy Analytics can be found here: https://www.strategyanalytics.com/access-services/automotive#.VuGdXfkrKUk