What’s completely misunderstood when discussing autonomous driving, is how far the gap has become between automakers and Waymo.
Long Tail, Long Nose, Short Nose
But lets start slowly. Take a look at some terms. Chris Anderson introduced the term Long Tail in 2004 in a Wired article, followed by a a popular book with the same title. An online retailer like Amazon sells a lot of copies from a small number of books, and few copies from millions of books. Individually, each book doesn’t sell much, but all together account for a significant amount.
Bill Buxton from Microsoft introduced the term Long Nose, in reference to the Long Tail. The Long Nose describes how a new technology often is under the radar for a long period, until it suddenly surfaces and wins the game. Knowing that allows Foresight-practitioners to discover trends earlier.
I wanted to discuss a special form of an exponential curve, that I call the Short Nose. We know exponential forms as they are seen at microprocessor development. Every 18 to 24 months, as Gordon Moore noticed in the 1960s, the number of transistors on a processor doubles, and therefore its speed. At the beginning the progress seems slow, but at a certain point in time it seems to take off.
The curve of a Short Nose is exponentially as well, but it goes seemingly fast at the beginning and the it seems to slow down.
The Short Nose of Autonomous Cars
Lets investigate the development of autonomous vehicles. Initially, when developers begin to define driving and traffic situations that autonomous vehicles should be able to master, you start with driving on a road or highway between the lane markers. Holding distance behind a car, changing lanes, exiting a highway through a ramp, right turns at an intersection, stop in front of a cross walk. Each of those scenarios can come in multiple variants. Add opposite traffic, and pedestrians. Then street signs, signal lights, construction cones, bicyclists, police cars, school buses, trucks, objects that fell on the road, and everything becomes much more complex.
From a few dozen situations we quickly realize, that there are thousands or even hundreds of thousands. Just add different light conditions according to the time of the day, weather conditions including direct sunlight, snow, misty fog, wind, or special events such as celebrations after a game, carnival, children at Halloween, hooligans, and the number of scenarios just explodes. Billions of fringe scenarios, many of them very rarely happening, are possible.
Not all combinations will happen (children at Halloween with hail), or are equally possible. Some scenarios you simply can’t invent, such as the elderly lady in a wheel chair driving in circles on a street chasing a duck (see video), or the half-naked guy running across the street and stopping to jump up and down the hood of the car. Others you ca imagine, but they are unlikely to ever happen.
The probability that many of those moments are going to happen is low. But we still want to prepare self-driving cars for such situation, so that they can react safely, in case they encounter that situation. To accomplish this, we have to test the cars in such situations.
Car makers that want to have their cars experiences those situations are recommended to have many of their cars drive a lot in real traffic. The probability that they encounter those situations increases.
In our example this means that with a handful of cars driving we can relatively quickly reach – in this hypothetically calculated example for illustration purposes – a safety level of 50 percent; of course only if we assume that we get the improvements put into the software. As we can see in the following chart, we can reach that with approximately 1,700 situations that our handful of cars have to encounter. But to reach a safety level of 60 percent, we need 7,700 covered traffic situations.
With each jump of safety by ten percent point, the number of required scenarios increases exponential. To get from 50 to 60 percent does not mean just 10 percent point more scenarios, but 450 percent more. Not 20 percent, but 450 percent safer is such a car on a safety level of 60 percent in comparison to one with a safety level of 50 percent.
But the effort increases dramatically, and to reach a higher safety level, the progress seems to become slower and slower. In our example we’d need more than five million covered traffic scenarios to reach 90 percent.
A company that’s not just using a handful, but hundreds of cars at the same time for testing, increases safety faster. One with thousands of cars on the roads of course even more fast. As some scenarios are occurring only rarely, maybe once a year in the test area, or only every five years, it can take some time until those are encountered and covered. The longer and the more intensive with many cars testing happens, and the larger the test area, the faster safety and capabilities improve for all cars.
A difference of 10 percent points from 50 to 60 could indicate that two years of testing were necessary to reach that. But the same 10 percent points between 80 and 90 can indicate ten years of development time. Even a gap of two years can have a regulatory agency come to the conclusion that companies with a low safety level on autonomous driving have to stop development and license the technology, to ensure the safety of the general public, and give the leading companies the license to operate.
Short Nose Versus Linear Development
How does the Shirt Nose differ from the linear development with lets say a Diesel? Lets assume that a combustion engine can be improved by two percent each year. That could be emissions, efficiency, speed, use, wear and tear, or a combination of all. This would mean that sooner or later such a combustion engine wouldn’t need much fuel, would be very quiet, wouldn’t wear out, and has the highest possible efficiency – until the physical and thermodynamic limits are met.
Each percent point progress in incremental innovation is bought with higher costs. The more the engine is optimized, the faster the costs go up for each additional percent point improvement. While the progress is linear, the costs are rising exponentially. Further improvement of the combustion engine quickly becomes uneconomical. Unless there is some new technology coming that allows a quantum leap. And that hasn’t really happened in a hundred years.
The best engine, as soon as it’s built into the car and shipped to customers, won’t improve. An improvement found after the delivery gets into the next generation of cars, but not into the ones already on the roads. That’s where the linear development, the incremental improvement stops.
An autonomous driving system on the other hand is getting better and better. The more cars are equipped with it, the more cars are encountering traffic situations, which then get added to the software, from which every car in the fleet benefits. This is a self-reinforcing effect, as we know it from networks.
One telephone is worthless. Two telephones can call each other and now have value. A third telephone allows me to call two participants, the value of my telephone has doubled. Each additional phone added to the network increases the value of each telephone.
Each additional car in an autonomous car fleet increases the fleet value, as the probability increases that the overall fleet encounters more traffic scenarios, and thus improves overall safety faster. The safer those cars, the more they are being used, the more they drive, the more traffic scenarios they cover, the safer they become. A self-reinforcing effect.
The Short Nose-example with autonomous cars has been only looking at the progress with the criteria discussed, and assumes that only the number of driven miles and cars used is important for improving safety. But there are phases where the algorithms and technologies used are not sufficient. Those roadblocks can only be overcome with new algorithms that have to be found, new sensors that have to be developed, or faster microprocessors.
Waymo And Competitors
Back in 2009 Waymo started first test drives of autonomous cars, known as project Chauffeur, and the as one of the Google X projects. Waymo so far has driven 9 million miles in autonomous mode. And the quality of those miles, of which Waymo adds 25,000 every day, is much higher. They are not collected by driving on relatively easy scenarios on the highway, but in challenging city traffic
So far Waymo has certainly encountered the simplest scenarios thousands of times in real traffic, and many many rare scenarios a few times. The 600 cars that Waymo operates help with that collection. And in the next months and years Waymo plans to bring an additional 82,000 cars on the streets. That’s by a factor of 136 more cars than Waymo operates today, and would at the same driving rate add 3.4 million miles a day to the tally. In three days Waymo would add more autonomous miles than in the entire 9 years since it started.
The only other company of which we know that has reached the million miles is Uber. But Uber stopped all test drives after a fatal crash earlier this year. GMCruise is another competitor that must have reached a higher number of miles driven. This company operates a fleet of over 100 cars.
The big unknown is Tesla. The situation is unique insofar, as Tesla has more than 200,000 vehicles on the road that are equipped with Autopilot Hardware Kit that is passively collecting data. Even that you can’t compare the data quality with Waymo’s, the sheer amount of data could compensate part of that gap. I still would count Tesla in.
Compare that with German manufacturers. There is the widely believed opinion that their gap to Waymo is around two years. As discussed and described above, nothing could me wronger. It’s a much wider gap that, if Waymo fully executes on their plans, is almost insurmountable.
And, I only talked about driving in real traffic. Waymo in addition operates an extensive simulation program, that increased the number of miles by a factor of 1,000. Each scenario of interest is simulated in thousands of variations. The simulator contributes to 80 percent of the improvements.
Those are all building blocks and efforts that, even if you have the money, cannot quickly build up, not to mention bridge the gap to Waymo. Like the universe is expanding ever faster, the same happens with the gap between Waymo and other vendors.
It’s no surprise that on the one hand investment bank UBS predicts that Waymo in 2030 will own more than 60 percent of the market for autonomous cars, and on the other Morgan Stanley sets Waymo’s value at 175 billion dollars. For everyone else there will be left either niches, or a fate like Microsoft’s Bing. Or they license Waymo-technology.
The hope of German automotive companies to leap in front with one of their magic tricks is unlikely. Waymo’s Short Nose shows them the long nose.
This article has also been published in German.