Back in 2007, I was a core member of the MIT DARPA Challenge team. We built one of the first self-driving cars to operate in urban environments. This was the pre-deep-learning era. It was fun, but we knew the road was going to be long. Fifteen years later, the dust is starting to settle and the time has come to take a candid look at the landscape. This article is a fairly opinionated perspective on the future of self-driving.
First, we will solve self driving eventually. It sounds like an obvious statement, but the latest news may have given cold feet to many people, and for a good reason. However, the stakes are simply too high to give up. Human, economic, sociological, and environmental benefits will be immense. Therefore, humankind will keep pouring billions of dollars and decades of engineering work into this problem until we solve it. Period.
Despite major progress in road infrastructure, car design and social behaviors, the casualties due to human driving remains unacceptably high. Even worse, the death rate per capita has been stagnating since 2010 . There is no promising solution to this problem other than taking the wheel off the hands of the humans (how do we solve drunk driving at scale?)
Self-driving cars will bring a revolution to our society. Not just in terms of how we simply move around, but how we live. Entire cities are built around cars. The very layout of our world and the fabric of our society will change forever.
It will take a very long time. For two reasons:
- The mileage death rate of human drivers stands at 1.46 deaths every 100 million miles driven . Self-driving cars will be broadly accepted only when they are orders of magnitude less lethal than humans. This means that we need an immense amount of data (and time) to demonstrate self-driving effectiveness.
- More importantly, we haven't cracked the AI problem
It will be a smooth transition. We all dream of a sudden transition with self-driving cars popping up everywhere all at once. This won't happen. Instead, we will go through a smooth transition that has already started.
The current generation of AI does not address self-driving. We think we are very close and keep talking about the remaining 1% of "edge cases". This vision is fallacious.
- The behavior on the 99% is still very brittle.
- The remaining 1% cannot be solved without solving high-level, abstract reasoning, which current AI technology cannot do
- It is unclear that deep learning can solve the problem single-handedly. We likely need completely new AI approaches to solve self-driving.
There exists an alternative where sensors embedded in the road network (or inter-car communication) provide hard robustness guarantees. Because we are talking about a major piece of infrastructure of the world and lives are at stake, hardwired robustness guarantees will be required much like they exist in elevators today.
Cybersecurity challenges are paramount. Once a country has shifted towards full self-driving technology, it is now exposed to another country or entity hacking its infrastructure with potentially disastrous consequences. This will play in favor of hardware-based guarantees.
We are over-estimating the ethical concerns.
True, some behavioral issues seem unsolved. If we know that a self-driving car will always be programmed to never run over a pedestrian, this opens the door to people jumping in front of cars for fun. In practice, people won't do this for the same reason they don't do a lot of other stupid things today. Cameras onboard the car will help solve legal issues, which will then become hardwired in the society through jurisprudence and social norms.
Some argue that self-driving cars cannot be solved without them being able to make moral decisions, having a unified moral standard for machines, and so on. This view is misguided. The social, economic and human benefits of self-driving cars will dwarf these concerns. For sure, regulations will be put in place and the legal responsibilities of self-driving car makers will be guaranteed, much like they are for plane makers today.
The intermediate milestone of AI-based driving assistance is useful but a dead-end for full self-driving. Lane departure detection, traffic jam driving assistance and the like will address major immediate frustration. These features are welcome and will make our lives easier in the short term. However, this branch is a dead-end with respect to full self-driving.
What problem are we trying to solve again? We need to constantly remind ourselves of the job to be done. We need to go from any point A to any point B, fast, safely, and on-demand. Self-driving cars turn out to be the only solution to this problem, but they will be part of a broader ecosystem (think: car as a mobility platform).