NHTSA’s Vision for Safety Severely Blurred
Autonomous driving systems are thundering down the road toward us. NHTSA is supposed to save us, but what are they doing? Their vision, in a document that is up for discussion, fails to understand automated cars and drivers. To begin with, they talk about 6 levels of automation and start with zero automation. Is that a horse? What are they talking about? They distinguish this from some level of driver assist: could they mean power steering? An automobile is already a huge driver assist system. That’s its benefit over a horse. Let’s get real and talk about the things that really matter, their so-called levels 4 and 5, where drivers can decide whether they want the computers to take over or not, and drive in all weather conditions or just some. If we want vision, then we need to focus on what matters, not rear view cameras, side radars, and speed controls: those are here to stay. Let’s not get hung up about not driving where construction starts, or through sleet and snow. Horses can go on backwoods trails. Most cars can’t. We know the limitations and are happy to drive on silky smooth highways.
All the auto manufacturers, with Google/Waymo leading the way, are already easily at level 4. That’s where the action is and where the confusion begins. Safer car drivers are the goal, so we ought to look at what makes a good safe driver, including autonomous ones, and especially focus on all the tips we give our sons and daughters to drive safely. Number one, for me, when they got over the basics was “go with the flow!” I have learned the hard way that cops don’t often give you a ticket if you’re in a flow of traffic, or if you stay under about 10% more than speed limits. How are they going to treat autonomous cars? Probably the same, but that would be a travesty, because driverless cars can probably lead the pack safely. We need a new look at speed limits for these cars if we are going to reap their benefits.
How risky is our driving? As an expert, I would say, in a new sedan with many driver assist features, your average risk is about a fatality in 100 million miles, give or take a few (it ranges from half that in Massachusetts to twice that in the Carolinas). Figure out how many years you have to drive and how many miles a year you drive, and you have an estimate of how likely you are to die in your lifetime. Let’s say you drive only 20,000 miles per year and you are going to drive another 50 years: that’s a million miles, and your chance of dying is 1 in 100 in your lifetime. That’s a lot lower than dying of cancer or many diseases we worry about a lot more than driving (700,000 per year or a chance of 10 in 100 in your lifetime) but still pretty significant. You should be happy if autos can reduce that.
Are these self driving cars (let’s just call them “autos”) the new chauffeurs? Probably a good enough metaphor to start with, with the caveat, that they can be better drivers than any human chauffeur. Humans make more than 90% of all the errors that lead to accidents. Autos may be able to get rid of most of these. And then there are the overly tired and drunk drivers: get rid of them and you could knock about 50% off the fatality rate. So think of autos as super chauffeurs with super powers. Batman would love them. They may not be the ultimate drivers for a few years yet, but all you have to do is read accident reports to see how stupid many of us are already; not very expert. I’m ready to part ways with human drivers for good. Let’s get licensed driving tests that really test your skills and only let those who know what they are doing drive themselves on smart lanes with 250 mph limits. The rest, the majority, will have to get an auto.
What about driving licenses for these new autos? NHTSA has traditionally stayed out of this realm, but examining their performance should be little different from testing cars for stopping distances or crash safety. While machine learning algorithms, neural networks, or deep convolutions sound impossible to assess, they are just computer software and code. What’s more they can be tested in simulators or real life track testbeds that mimic the full range of pedestrian, weather and construction road hazards. NHTSA should get some testbeds ready and start giving these current level 4 drivers some realistic tests too. This does not mean setting regulatory standards yet; but, when accidents start to happen with different frequencies for different manufacturers, NHTSA will need to be ready before it is too late.
It is a copout to worry about ethical decisions, such as whether to run over a pedestrian or crash into a schoolbus full of kids, since most current drivers have no idea which they would do in an emergency; but it is a much more serious issue whether to obey legal speed limits punctiliously to the letter or drive safely by staying with the flow of traffic. NHTSA needs to assert that it will examine how autos begin to deal with this issue, and reveal at the beginning that safety dominates legalistic frameworks, but some reconciliation of the issues needs to be worked out. One potential direction is for autos to compute the average of nearby cars, and be allowed to drive at some compromise between the speed limit and the average. This would include merging safely into the ongoing flow of traffic from turns or lane changes. Safety first!