I just finished listening to Autonomy by Lawrence Burns, a former VP of R&D at GM exec and a (former?) strategy consultant for Waymo. The focus of the book is the history of self-driving car R&D and the potential future impact on society. I’m disappointed that there isn’t more in the book about the actual technology.
Topics that are emphasized and repeatedly discussed in the book:
- HD maps
Topics that are not discussed in the book:
- neural networks
- computer vision
- machine learning, deep learning, reinforcement learning, imitation learning
At one point, the author seems to misdescribe Mobileye’s computer vision technology as comparing objects on the road to a large library of images. This sounds like a garbled description of a neural network. Later, there is a brief mention of HW1 Autopilot — using Mobileye’s EyeQ chip — being taught to recognize vehicles, but it isn’t explained beyond that. (Edit: The Mobileye chip that did vision for HW1 Autopilot might not have used a neural network, so Burns’ description might be accurate. But, still, if this book were all you had to go on, you would not be aware of the existence of neural networks.)
The basic concept of a neural network or machine learning is not explained in the book. It only goes as far as to describe the “software” or “code” or “programming” of the vehicles, giving the impression that everything in a self-driving car is just conventional, hand-coded software. That hand-coded software is also only discussed in hazy, general terms.
The way I see it, autonomous cars would not be possible without machine learning (at a minimum for perception), yet from reading the book you wouldn’t even know machine learning exists. You would think autonomous cars are just a matter of HD maps, lidar, and hand coding.
If you want to learn about the internal politics at Waymo, or hear stories from the DARPA Grand Challenge, then you might like this book. If you want to learn about the technology, I would recommend Lex Fridman’s MIT course instead.