California requires all companies that test self-driving cars on state roads to provide kilometers and the number of “overrides” or times when a human driver takes control. Kyle Vogt, co-founder and CTO of Cruise, believes that this reporting method is a poor metric for comparing companies and causes companies to run tests and demos in simpler environments to reduce reported crashes.
In a post on Medium, Vogt says that when cruising, disengagement is sometimes used as a favor for other drivers or as a cautious reaction by the driver to a situation that the vehicle could have dealt with. He explains: “Have you ever been in the back seat of a human-powered car and needed to grab the steering wheel when something crazy happened on the street? It is exactly like that. ‘
Autonomous vehicle manufacturers (AV companies) must be particularly careful when it comes to security and perception of security. AVs that work correctly are not messages, but decommissioning, red lights, and crashes are very many messages that can affect your perception of AV security, regardless of your actual security record. This leads to the well-controlled demos with which Vogt has problems.
“Companies carefully curate demo routes, avoid urban areas with cyclists and pedestrians, restrict geofences and pickup / drop-off points, and limit the types of maneuvers that the AV performs while driving – all to limit the number of disconnections. After all, an AV is only ready at prime time when it can do dozens, hundreds, or even thousands of such trips without anyone touching the steering wheel. It’s the ultimate sign that technology is ready, isn’t it? Not correct.”
According to Vogt, Cruise will release data that will improve AV performance once the vehicles are ready for use. This includes: “a) data on the actual performance of human drivers and AV devices in a specific environment and b) an objective comparison between apples with statistically significant results.”
I look forward to this information and hope that it will convey AV security, which is often falsified and obscured by incomplete data, more clearly. I am sure that autonomous vehicles will one day be a great thing for overall security and accessibility. To know when this time is reached, we need a clearer picture of AV security and comparability with human drivers in real environments.