Skip to main content

Book ,

Contemplating the path forward for autonomous vehicles

Driverless: Intelligent Cars and the Road Ahead

Hod Lipson, Melba Kurman
MIT Press
Purchase this item now

Everyone is talking about driverless cars. Recent advancements in power storage, computational power, sensory technology, communication bandwidth, algorithm efficiency, and more have moved autonomous vehicles from the realm of science fiction into the real world. In Driverless, authors Hod Lipson and Melba Kurman delve into these and other contributing technologies.

Included is an exhaustive review of one such advance, “deep learning,” that allows a computer to learn from experience. One boggling potential application in the context of driverless cars is “hive learning,” in which everything a vehicle learns about road conditions can theoretically be instantaneously shared with all other vehicles, enabling a new vehicle to possess the experience of millions of miles of driving. Yet, deep learning is only as comprehensive as its training. The deadly Tesla accident that occurred in May seems to have resulted from an encounter with a “corner case,” a condition not included in training.

Humans are vulnerable to corner cases, too. A driver who faces a new situation without preparation is experiencing a corner case. And human information sharing is not instantaneous, meaning that any lessons learned will benefit only the individual driver.

Unlike traditional computer algorithms, deep learning creates a nonexplicit, black-box intelligence that cannot be reverse-engineered. For this reason, trying to determine why a driverless system made a decision can be as difficult as ascertaining a human’s decision process. This, write the authors, will have implications for liability assessment.

The book also delves into advances in machine vision. Only recently have we developed pattern-recognition software and hardware capable of recognizing real road challenges such as shadows. The Tesla accident, which appears to have involved difficulty discerning the side of a turning tractor-trailer, highlights that there is still work to do.

“Human-in-the-loop” computation automates most, but not all, functions. It is the approach used by pilots and surgeons. But the authors contend that drivers will be unable to maintain sufficient alertness over the course of a lengthy trip. It was reported, for example, that the driver of the Tesla was watching a movie at the time of the fatal accident.

The authors acknowledge that an autonomous future is not guaranteed and discuss potential problems, both in technology and in policy. Yet they are clearly rooting for it. After reading this book, you will be knowledgeable enough to make your own informed opinion.

About the author

The reviewer works in software reliability and accident reconstruction for Engineering Systems, Inc., Ann Arbor, MI 48108, USA.