Abstract: a Cruise examine suggests they’re higher than a younger male journey hail driver in a leased car. Nonetheless, this result’s an estimate, as a result of there may be not but sufficient knowledge to have a agency conclusion.
I’m glad to see Cruise launch a paper describing the methodology for computing the human driver baseline, which that they had not beforehand finished. Identical too for his or her “significant threat of damage” estimation methodology. And it’s good to see a benchmark that’s particular to a deployment fairly than a US common.
Cruise has revealed a baseline examine for his or her security evaluation right here:
weblog submit: https://getcruise.com/information/weblog/2023/human-ridehail-crash-rate-benchmark/
baseline examine: https://deepblue.lib.umich.edu/deal with/2027.42/178179
(notice that the baseline examine is a white paper and never a peer reviewed publication)
The necessary take-aways from this by way of their robotaxi security evaluation are:
- The baseline is leased journey hail autos, not odd privately owned autos
- The drivers of the baseline are younger males (nearly a 3rd are beneath 30 years outdated)
- A “significant threat of damage” threshold is outlined, however considerably
arbitrary. They apparently would not have sufficient knowledge to measure precise
damage charges with statistical confidence. Provided that we’ve got seen two
accidents to Cruise passengers to date (and no less than one different damage
crash), this isn’t a hypothetical concern.
It must be no shock if younger males driving leased autos as
Uber/Lyft drivers have a better crash charge than different autos. That’s
their baseline comparability. In equity, if their enterprise mannequin is to
put all of the Uber and Lyft drivers out of labor, maybe that may be a helpful
baseline. But it surely doesn’t scale to the final driving inhabitants.
A conclusion {that a} Cruise robotaxi is safer (fewer accidents/fatalities)
than an odd human driver will not be fairly supported by this examine.
- It isn’t an “common” human driver except you solely care about Uber/Lyft. If that’s the concern, then OK, sure, that may be a affordable comparability baseline.
- I didn’t see management for climate, time of day, congestion, and different circumstances within the baseline. Highway kind and geo-fence had been the points of ODD getting used.
- There’s inadequate knowledge to have a conclusion about damage charges, though that may come considerably quickly
- We’re a great distance away from perception into how fatality charges will flip
out, for the reason that examine and Cruise have about 5 million miles and San
Francisco fatality charge is extra like one per 100 million miles - The Cruise emphasis on “at fault” crashes is a
distraction from crash outcomes that should essentially embody the
contribution of defensive driving conduct (avoiding not-at-fault
crashes)
This examine may assist a Cruise assertion that they’re on monitor to being protected based on their chosen standards. However we nonetheless
do not know the way that may end up. This isn’t the identical as a declare of confirmed
security by way of hurt discount.
A special report doesn’t construct a mannequin and estimate, however fairly
compares precise crash stories for robotaxis with crash stories for journey
hail automobiles and involves the conclusion that Cruise and Waymo operated at 4
to eight occasions as many crashes as common US drivers, however that their crash
charge is similar to journey hail autos in California.
https://www.researchgate.web/publication/373698259_Assessing_Readiness_of_Self-Driving_Vehicles
