Moore’s Law for Autonomous Driving

Although the Christmas holiday is approaching, the autonomous driving industry is still "hot".

Recently, May Mobility, an autonomous driving company from the United States, successfully finalized a $50 million Series B financing round from the super giant Toyota.

The funds injected by Toyota will be mainly used for company expansion, such as adding shuttle vehicles, recruiting engineering and operations personnel, etc.

May Mobility has currently deployed autonomous shuttle services in three US cities (Detroit, Providence, and Grand Rapids), with a fleet of 25 low-speed shuttles.

Less than a year ago, May Mobility successfully obtained US$22 million in Series A financing.

Looking ahead, May Mobility hopes to increase the number of shuttles in the three cities to 25. Such a fleet size will not only help the company generate more revenue, but also have a real impact on the traffic in the city.

For May Mobility, this round of financing means more than just filling its pockets. The company, which was established in 2017, finally has the endorsement of Toyota, a super giant.

It is reported that Toyota has included May Mobility in the company's list of "future open platform autonomous driving suppliers."

Although neither Toyota nor May Mobility has shared details of their collaboration, industry insiders believe that the Japanese giant is exploring autonomous driving solutions for its own e-Palette platform (debuted at CES 2018).

Although not as famous as Aurora, May Mobility is also a "Big Three" company, with co-founders including: Alisyn Malek (COO of the new company), Edwin Olson (CEO of the new company) and Steve Vozar (CTO of the new company).

As the company's CEO, Edwin Olson has been deeply involved in the autonomous driving industry for more than ten years.

When he was an assistant professor at the University of Michigan, he participated in the 2007 DARPA Urban Challenge, and later served as the chief researcher of Ford's autonomous driving project and the deputy director of autonomous driving research and development of Toyota Research. He even summarized a set of Moore's Laws applicable to autonomous driving.

Next, let’s follow Olson’s article to learn more about what Moore’s Law of autonomous driving is.

Edwin Olson's Moore's Law for Autonomous Driving

As the CEO of an autonomous driving company, I often ask myself, when will self-driving taxis become mainstream?

From an industry perspective, people’s understanding of this issue varies greatly. A company that is good at doing business gave a node in 2019, while engineers think it is still quite far away. So who should we listen to?

In the following, we measure the performance of the autonomous driving system based on the takeover rate.

The so-called takeover rate here actually refers to the frequency of intervention by the safety driver. For the same mileage, the lower the frequency of takeover, the stronger the performance of the autonomous driving system will be.

Moore's Law?

Intel founder Gordon Moore proposed the concept of "Moore's Law" to reveal the speed of progress in information technology, that is, when the price remains unchanged, the number of components that can be accommodated on an integrated circuit will double every 18 months, and the performance will also double.

In other words, the computer performance that can be purchased per dollar doubles every 18 months.

This exponential growth also means that today’s smartphones can outperform the supercomputers of the past.

Of course, exponential growth is rare.

Take trees and people for example, their growth is linear, so it is much slower.

Most things that grow exponentially don't last long enough. Bacterial proliferation is a prime example; once their living space becomes crowded, their exponential growth will come to a halt.

In fact, the current improvement in computer performance can hardly meet Moore's Law.

In other words, it is not uncommon for a newly emerging technology to grow exponentially. Although this is a relatively optimistic assumption, it is very consistent with the expectations of those who want to change the world with technology.

So today we might as well speak our minds and make bold assumptions to see what the future of self-driving cars will look like.

Here, we assume that autonomous driving technology will experience exponential growth.

In other words, this article is about weaving a Moore's Law for autonomous driving. However, the result may not be what you want.

data

In 2004, the best self-driving car in the world was undoubtedly Carnegie Mellon's "Sandstorm", which "won first place" in the first DARPA self-driving challenge, even though it only ran 7.4 miles of the 150-mile race.

There is absolutely no intention to mock Sandstorm here, after all, other models of the era performed worse.

To summarize, in 2004, the takeover rate for autonomous vehicles was about 10 miles per trip.

14 years later, Waymo has increased its takeover rate to 11,017 miles per trip (104), which is definitely an exponential improvement.

With these two numbers, we can calculate Moore's Law for self-driving cars, which states that the distance traveled between takeovers will double approximately every 16 months.

Did you find anything?

Moore's Law for self-driving cars is almost the same as it is for the computer industry - performance doubles every 16 months.

The black line in the above figure represents the development of self-driving cars from 2004 to 2018.

If we continue to extrapolate (the red line), one day we will intersect the blue line, which represents human driving capabilities.

Note that the Y-axis here uses a logarithmic scale, so exponential growth looks like a straight line.

The key question here is: "How good does an autonomous driving system have to be?"

Assuming our goal is to match human driving ability, we would need to achieve a probability of one accident per 100 million miles (10⁸).

Does it feel a bit abstract?

In fact, if I explain it to you, you will understand that the average person may only drive a car for hundreds of thousands of miles in his or her lifetime.

As for self-driving cars, the cumulative mileage of all manufacturers' test cars since their inception is probably no more than 20 million miles.

By comparing these results, you will find that humans are indeed much more powerful now. After all, the difference between 10⁸ and 10⁴ is 10,000 times. In other words, the level of self-driving cars is only 0.01% of that of human drivers.

Even if the trend of doubling performance every 16 months can be maintained, it will take another 16 years (2035) to reach human driving levels.

Obviously, those claims that "autonomous driving will be achieved this year or next year" are not accurate.

Although manufacturers will still showcase their latest technologies in a high-profile manner, this does not mean that their systems can compete with human drivers.

Of course, most autonomous driving failures are not fatal, and if the constraint is changed to only cause injury, humans can safely drive up to 10⁷ miles.

So, if the failure of self-driving cars is not fatal, these companies can get human drivers' licenses four years earlier than they would otherwise have to wait until 2031.

The evolution of autonomous driving

May Mobility is developing self-driving cars, and our products are already on the road.

You might say, this is not far from 2035, how did you do it?

There are two main reasons:

May Mobility does not intend to compete with human drivers in all scenarios. We focus on a small subset of all driving tasks, namely low-speed (< 25 mph, about 40 km/h) autonomous driving on fixed routes.

Simpler routes combined with lower speeds mean a significant reduction in overall complexity.

At the basic level, May Mobility uses different technologies to help vehicles understand what other vehicles and pedestrians are doing.

We believe that this technology (multi-strategy decision-making technology) can enable May Mobility to take a different path and bring transformative impact to the entire industry.

There is precedent for this kind of change. In 2011, if you wanted to do a complete DNA sequence on someone, it would cost $100 million.

But in the following years, the technology's progress began to follow Moore's Law, with performance doubling every 20 months.

As can be seen from the above figure, the development of DNA sequencing technology followed Moore's Law between 2001 and 2007, with a relatively stable exponential growth rate during this period.

But in 2008, a new technology for gene sequencing was introduced, which completely changed the "coefficient" of Moore's Law for gene sequencing, and the progress curve became steeper. The performance that used to double in 20 months can now be doubled in 4 months.

This technological advancement has really put us on a rocket ship, bringing us decades closer to our goal of completing gene sequencing for $1,000.

We believe that similar strokes of genius will also occur in the field of autonomous driving in the future. In May Mobility's view, this revolutionary new technology is its own multi-strategy decision-making technology.

Summary of key points:

• 
  

• 1. The performance of self-driving cars doubles every 16 months, which is Moore's Law.

• 
  

• 2. Given that the level of self-driving cars is only 0.01% of that of human drivers, it would be a daydream to talk about self-driving taxis before 2035.

• 
  

• 3. Moore's Law for self-driving cars also has two loopholes; one is that revolutionary technology may greatly change the growth curve; the other is that some companies may abandon the "all-knowing and all-powerful" all-weather autonomous driving and instead choose simpler application areas.

• 
  

• 4. Moore's Law of autonomous driving will make autonomous taxi companies cry, but it will definitely be good for shuttle companies.

If the above brief description is not enough, the following three points must not be missed.

Point 1:

Moore’s Law for self-driving cars depends on the data we use. If you think that self-driving companies are overestimating their technology, then the timeline for the deployment of robotaxis will be pushed back again.