Two approaches to autonomy.

I think it’s fair to say that at a high-level of abstraction, there are two types of approaches to building self-driving cars.

I have been calling the approach that most companies use the “robotics” approach, or the “traditional” approach, or the “classical” approach. I just read a paper calling this approach “mediated perception”, a term I’d never seen before.

This approach combines computer vision, sensor fusion, localization, control theory, and path planning.

Unfortunately, the terms “traditional” or “classical” make this approach sound outdated— which it isn’t, in fact this is the state of the art. And, so far as I know, nobody uses the term “robotics” to describe this approach, except me.

The other, second approach is currently in vogue with Silicon Valley startups, and involves training a single deep neural network to take sensor inputs and produce steering, throttle, and brake outputs. This approach is sometimes called “behavioral cloning”, or “end-to-end driving”, or “deep learning”, or I just read a paper calling it “behavioral reflex”.

So we have a little bit of an issue with naming conventions. It would be nice if we could just use one term to refer to each approach, and maybe there are settled terms and I’m just missing them.

Fill me in if you know what I’m missing.