Billions are being invested in self-driving technology each year, and some question why it isn’t further along today, whether it can be really made to work and whether it can be a business that scales and makes profits for the many companies in the game.
To address these questions, I engaged this week in a formal debate. Taking the negative was Professor Raj Rajkumar from CMU. Raj was one of the team leaders when CMU won the DARPA Urban Challenge — the contest that really got the self-driving robocar world going.
For those who prefer, here is a transcript of my opening statement. For the rest you will need to go to the debate. In addition to the opening and rebuttals, there are 3 sub sections on:
- Is Tesla’s approach viable at all?
- What are the prospects for self-driving consumer vehicles?
- What are the prospects for commercial services.
I’m going to outline why self-driving vehicles are well on their way to being mainstream in 10 years. To me that means they are available in 50 or more major cities, and being sold in decent volumes to private individuals for freeways and major roads. That doesn’t mean you can buy a private car that does minor city streets, or get a ride in every town.
Obviously, we now see pilot services operating in 31 cities around the world like San Francisco, Phoenix, Austin, Shanghai, Beijing, Shenzhen, Wuhan, Toronto and many others, with smaller delivery robots in many more. The passenger services are pilots, though some of the delivery services are production. I’ve made a map of cities where services have been or shortly will be deployed with no employee in the vehicle and it’s much more populated than people expect.
This is coming through the march of Moore’s law combined with the greatest effort in robotics in history, with a huge amount of detail work.
To understand just how this grows, you must understand that the core target is robotaxi, not private car, and that perfection is not necessary — it’s not even the right goal — before deployment.
What matters to deploy a taxi or delivery service is whether it’s commercially viable. It may only handle a particular region or subset of streets, and may be supplemented with carshare, transit, TNCs and car delivery. It doesn’t have to do everything. To get really big, it has to offer car replacement to the urban dweller, but it can be quite lucrative with the baby step of replacing Uber.
You can do this when you’ve reached a suitable level of safety, as well as of road citizenship. To do that you need an “MTBF” better than human. You need your error rate low enough, and the severity of errors low enough but you don’t need to be error free.
To solve this, one basic approach is just to have seen so much in millions of miles of testing and billions of miles of sim so that the frequency of surprises is low. Even though corner cases don’t just happen at the corners, you can measure how often you’re surprised. And you can also measure how often you handle surprises well and how often you fail, and from that see if your MTBF is low enough. Multiple companies have already reached that level in their pilot projects, at least on safety, and they’re working rapidly on road citizenship.
They’ve made a raft of great tools to get this confidence. Amazing LIDARS with great resolution and performance that will never miss that something is there. Imaging radars with velocity detection and weather penetration. Even though CV isn’t ready for prime time on its own, its improving rapidly, improving the most important task of prediction and dropping in price. All these sensors are plummeting in price and soaring in performance. If you don’t believe in them today it would be foolish to bet against them a few years from now.
The leader is Waymo. I worked on that team a decade ago and continue to be impressed. Cruise is behind but also reaching milestones. We’ve all seen the press over both teams having stall-outs and little mistakes here and there, blocking traffic, being confused by unusual situations, or waiting for a rescue driver when they team doesn’t trust remote assistance to be 100% safe. These may seem silly but it’s exactly what you want to see — the cars on on the roads to find mistakes and fix them.
Nobody’s been hurt by Waymo and Cruise’s incidents remain minor and better than human. The great thing about robots is, unlike people, they don’t repeat the same mistake twice. Almost every problem you see is a success — something found that won’t happen again.
For skeptics, it’s also worth nothing that communications tech also is booming, and nobody doubts that long before 10 years we’ll see highly reliable, fast and low latency communications in all the needed zones, enough so that if you had to, problems could be solved with remote driving, though that’s not how anybody wants to do it — but they will if the option is to not deploy.
China is also going very strong. Over 300 teams are working on vehicles and there are dozens of fully autonomous deployments. Regulation will not stand in the way there, though it’s an open question in the USA, and definitely in Europe. Indeed, if I had to do this debate for Europe I would be less optimistic, though Mercedes promises their S-Class will do 130kph on the Autobahn without driver attention fairly soon.
While I’m not at all a fan of infrastructure based approaches, and they are doomed in the west because the pace of change and innovation in infrastructure is measured in decades, it’s not at that pace in China — just look at what’s happened in the last 10 years, and many Chinese teams are working on infrastructure based approaches.
To return to the core thesis — the reason people don’t think this technology will succeed is they try to define the problem as much bigger than it is. To get a real business serving lots of people you only need to do decent robotaxi service in lucrative sized zones of major cities. You don’t need to drive every road. We’ll talk later about consumer cars, but they don’t need to be that heavily featured either.
Considering autonomous vehicles are already on the road in over 30 cities, it’s a slam dunk to being a lot bigger by 2033, unless the law shuts it down, and right now it doesn’t seem on track to do that, except in a few places. When the change comes it will boggle your mind with its speed. This newspaper clipping from 1928 describes how the horse and buggy industry completely collapsed in 2 years when its time came. Moore’s law always surprises you — don’t bet against it.