In a stunning statement at a D.Live conference earlier this month, Waymo’s CEO, John Krafcik, said it would be decades before fully automated vehicles were ubiquitous. In other words, according to The Information, “Uber and Lyft won’t be threatened by self-driving cars anytime soon.”
Krafcik added that, “This is a very long journey. It’s a very challenging technology and we’re going to take our time.”
When asked by CNN if self-driving cars were closer than most people think to being road-ready, he replied, “I would say their first applications are close, ubiquity is further away than most people would believe.”
By first applications, he meant we would see cars that would be able to drive in small controlled areas of cities, presumably during the daytime and when weather permitted. (Weather is another big problem for these cars currently).
This is something many close observers have said for a long time. That despite all the media hype that these cars are just around the corner, they are actually a long ways off.
Why the Hold Up?
Actually, there is no hold up. The truth is, in my opinion, that these cars have been much farther away from completion than we’ve been led to believe. When we are shown video clips of driverless cars taking a few laps around a track or told amazing statistics like they’ve driven millions of miles without an accident, it makes it sound like they’re right around the corner. But I don’t think they are, or ever have been.
Making a fully automated vehicle is a lot more difficult than it appears. What we’re not usually told is that it is much simpler to program a car to drive itself around a well-mapped closed track than it is to design one that can drive anywhere and everywhere at any time day or night, rain sleet or snow.
It’s also much easier to build an autonomous vehicle that can travel several square miles of a well-mapped and well-planned area of a city at specific times during perfect weather conditions than it is to let the car loose to travel on its own wherever and whenever it likes. In other words it’s much easier to build the first 90 percent of the technology than it is to build the last 10 percent.
Anyone who has paid close attention to autonomous cars has noticed the numerous problems they have encountered. For instance, we often hear how many millions of miles Google’s Waymo cars have driven without an accident. But we’re rarely told how few miles they can go without human intervention.
That story isn’t quite as exciting as watching a car drive itself around a track! So, it doesn’t get reported nearly as often – which leaves us all with the misimpression that they’re right around the corner.
An Accident Waiting to Happen Every 13 to 6,000 Miles!
In 2017, Waymo’s cars reportedly drove 6,000 miles before human intervention was necessary. That sounds pretty impressive – and it is. It’s amazing humans have been able to develop a car that can drive up to 6,000 miles completely on its own!
But if you think about it for a minute, you’ll realize that while that is indeed a great achievement, it’s nowhere close to what we need for these cars to fully replace human drivers. Every time these cars require human intervention, the intervention is needed to avoid a possible accident. Without that human intervention we can assume an accident of some kind could have occurred.
To put that in perspective, imagine that you had one accident for every 6,000 miles you drove! Some would be minor mishaps perhaps but others would be serious. And they would happen every 6,000 miles! For many drivers, that would mean having a car accident every other month! And that’s truly unacceptable.
For full-time Uber drivers, who drive a ton of miles, they would be having accidents every two to three weeks! When you think about it like that you can see that as impressive as the 6,000-mile figure is, these cars are nowhere near being ready. Waymo’s cars are the best of the best and even they fall far behind the performance you would get with any human driver.
To make matters worse, the other autonomous car players are doing much worse. GM’s Cruise can only go about 1,500 miles without human intervention.
And to give you an even clearer picture of what’s going on in this industry, Uber’s self-driving cars – you know, the ones that will supposedly replace all Uber drivers within the next year or two – they require human intervention every 13 miles! Can you imagine having an accident every 13 miles? That would be like having three to four accidents every day for most drivers! Well, that’s Uber’s self-driving car.
Waymo’s cars are in fact 461 times more advanced than Uber’s. To put that number in perspective, let’s compare it to Uber’s surge pricing.
Say you take an Uber trip that would normally cost $15 but this particular trip happens to be at 12:40 a.m. on New Year’s Eve when it’s pouring down rain and everybody in town is calling for an Uber at the same time. The surge premium has skyrocketed, not to 2 or 3 or 4 times, but to 461 times the normal price. That $15 trip would cost $6,951! That’s the difference between the state of Uber’s self-driving program and Waymo’s.
These numbers are probably not even giving us the full picture. Waymo and the other companies test their autonomous vehicles all over the world. And in many places they test, disclosures aren’t mandatory. Additionally, all the testing they do on private property never has to be reported.
So, these numbers only account for test drives conducted on public roads in jurisdictions where they are required to report. If we had access to all the numbers, the results might be even worse.
Why is it So Difficult?
It’s very easy for us humans to underestimate the complexity involved in even the most simple tasks we perform. We look out of the windshield of our car and see a crumpled piece of paper on the road ahead of us and we immediately know that it’s not a rock. But a simple task like this s no easy feat for a computer. A self-driving car would swerve around a plastic bag as quickly as it would swerve around a rock!
Another big hurdle is the fact that Waymo’s cars are designed to use computerized maps of the areas they drive in. So, before a car can be set loose in an area, Waymo has to send out “learning” cars to drive the specific routes over and over again. The data is then taken in and analyzed by humans and computers before it is ready to be used in the vehicles.
But there’s just one little problem with this approach. If any changes are made on the route after the data is fed into the computers, the cars will not be able to recognize or deal with those changes. If, for instance, a city puts up a new stoplight at an intersection the day after the data is collected, the cars will not know that stoplight is there and they won’t stop for it.
It would be literally impossible to have these “learning” cars traveling and mapping every single mile of American roads every minute of every day looking for changes. So, this is another huge hurdle they are going to have to overcome before they’ll be ready to go.
And since self-driving cars see people as moving pixels, they can’t tell the difference between a woman crossing the street or a police officer waving frantically for all cars to keep moving.
Self-driving cars are suffering from the “last-mile” problem, which is where it has taken them ten years to get the cars to 85-90 percent proficiency, but it could take twice that long to finish the last 10 to 15 percent.
It’s the same problem you see in a lot of areas of technology. Like machine language translation. If you’ve used Google Translate over the last few years you’ve seen a vast improvement. However, it’s still not quiet there. It still spits out translations that are awkward at best and incomprehensible at worst.
Now, imagine that program is driving your car. Every time it spits out an awkward translation – that’s a car accident.
The first machine language translation was achieved in 1954. Researchers who worked on the project expressed their belief that machine translation would be a solved problem within three to five years! Well, here we are 65 years later, with vastly superior computer technology – and it’s still not completely solved.
Why is There So Much Hype if Self-Driving Cars Are Not Even Close to Ready?
Most of the hype comes from the media, which loves this story because it’s exciting and new and it makes for great click-bait. But it is also helped along a bit by companies like Uber who often publicly brag that they’re very close to making this a reality. Why they do that, when there’s so much work left to do before they are truly ready, we can’t know for sure. But an upcoming IPO might have something to do with it.
If investors believe this technology is just about ready for prime time it would increase their confidence that Uber would be a good investment, because it should cut their driver costs tremendously. And that would make them far more profitable than they are today. So, it could help their initial stock offering price if investors think the technology isn’t that far off.
What This Means for Rideshare Drivers
This means Uber and Lyft drivers who have come to rely on the income they make from driving – don’t have anything to worry about. At least probably not for 20-30 years or so. Not only do numerous technical details have to be ironed out but then you’re going to have a massive amount of regulatory problems that will have to be worked out, as well as insurance problems.
The autonomous car companies are either going to have to figure out a better more efficient way to do mapping or they’re going to have to ditch mapping altogether. Or possibly use a blend of mapping and coordination with every city and county in the nation which could feed them any and all road changes they make.
But, if you’re an Uber, Lyft or any other kind of driver who makes an income from driving, it looks like your simple human ability to drive a car is going to far surpass those of machines for decades to come.
(Updated January 15, 2020. Originally published 11/30/2018)