Humans are remarkably safe at driving
Here’s a quick question: how far would you have to drive before you die in a car crash? Answer out loud to yourself before you look it up.
The dirty secret in the autonomous industry is that humans are actually incredibly safe. Most humans in a matter of a few hours can operate any multi-ton vehicle in most weather anywhere in the world. As road conditions change humans instantaneously adapt control response and safety protocols. Humans dynamically interact with totally never before seen objects and situations. They regularly fall asleep while going over the speed limit. They regularly drive drunk, and still manage to avoid crashes with sub-second steering and brake responses. Also realize that there is a difference between how far one has to drive before there is a slight fender-bender versus a fatality (which is more?).
So, what is the number of miles one human would have to drive before a dying in a crash? Say it out loud. Understand that on average people drive about 10,000 miles a year.
What is your answer? The answer, and this includes teenagers and drunk drivers and people texting, is about 100 MILLION miles. A bit less or more in different countries or years, but about that. I’ve talked to Waymo engineers who didn’t realize this the first many years they worked on self-driving cars, and when you work on something without realizing how hard the problem is you don’t necessarily work on it from the best long term perspective.
As an industry, we need to do a different approach than one that hits a wall as does self-driving with highly trained drivers as Waymo, Lyft, and Uber are doing. We need to get the miles and data cheaply. Amazon delivery vehicles can drive slowly enough that it is a truly useful limited domain problem, yet can be largely automated, and already has a trained, paid supervising driver. The marginal cost of collecting this mileage training data is near 0. That’s a viable start. Maybe there are others like those of Tesla, and Comma. I know Waymo’s is a dead-end demo.
How did we verify that rockets or planes worked without flying 1e99 miles? It’s because of physical laws and design. The problem is that, surprisingly, rocket science is easier than self driving cars. It’s because physics is so mechanical. Landing on Mars is easy comparatively, since everything is stochastic and measurable, and not as chaotic or reflexive (maybe Martian wind is a bit chaotic). Every mile in space is almost exactly the same. Self-driving cars involve other humans on the road, and predicting them, each meter a different meter, no two the same. Much more complicated. The sensor and actuator failures are manageable using Physics eventually, but not Prediction (the hard part!).
If you liked this, then you’re going to love these short pieces Why Tesla Autopilot ought to be awful, until it’s perfect, and also the piece Tesla Saves Lives, Waymo Gambles Passenger Lives.
