Rage against the machine? Google's self-driving cars versus human drivers

Teoh, Eric R. / Kidd, David G.
Journal of Safety Research
August 2017

Abstract
Background: Automated driving represents both challenges and opportunities in highway safety. Google has been developing self-driving cars and testing them under employee supervision on public roads since 2009. These vehicles have been involved in several crashes, but it is of interest how this testing program compares to human drivers in terms of safety.
Methods: Google car crashes were coded by type and severity based on narratives released by Google. Crash rates per million vehicle miles traveled (VMT) were computed for crashes deemed severe enough to be reportable to police. These were compared with police-reported crash rates for human drivers. Crash types also were compared.
Results: Google cars had a much lower rate of police-reportable crashes per million VMT than human drivers in Mountain View, California, during 2009 to 2015 (2.19 vs. 6.06), but the difference was not statistically significant. The most common type of collision involving Google cars was when they got rear-ended by another (human-driven) vehicle. The Google car shared responsibility for only one crash.
Conclusions: These results suggest the Google self-driving car program, while a test program, is safer than conventional human-driven passenger vehicles; however, currently there is insufficient information to fully examine the extent to which disengagements affected these results.
Abstract Background: Automated driving represents both challenges and opportunities in highway safety. Google has been developing self-driving cars and testing them under employee supervision on public roads since 2009. These vehicles have been involved in several crashes, but it is of interest how this testing program compares to human drivers in terms of safety.
Methods: Google car crashes were coded by type and severity based on narratives released by Google. Crash rates per million vehicle miles traveled (VMT) were computed for crashes deemed severe enough to be reportable to police. These were compared with police-reported crash rates for human drivers. Crash types also were compared.
Results: Google cars had a much lower rate of police-reportable crashes per million VMT than human drivers in Mountain View, California, during 2009 to 2015 (2.19 vs. 6.06), but the difference was not statistically significant. The most common type of collision involving Google cars was when they got rear-ended by another (human-driven) vehicle. The Google car shared responsibility for only one crash.
Conclusions: These results suggest the Google self-driving car program, while a test program, is safer than conventional human-driven passenger vehicles; however, currently there is insufficient information to fully examine the extent to which disengagements affected these results.