A comparison to Waymo’s auto liability insurance claims at 25M miles

in the same weather?

*retracted :)

This is covered in the paper, although, somewhat ironically, opposite the bias you're implying:

>The garaging zip code of the insured vehicle was used as a proxy for the city (Phoenix, San Francisco, Los Angeles, Austin) in which the vehicle drives. Waymo also almost exclusively operated on surface streets (non access-controlled freeways) with a unique distribution of driving that is representative of a ride-hailing fleet. In contrast, the benchmark represents the privately insured driver population that resides in these geographic regions. The associated benchmark mileage has more freeway driving than the Waymo ADS. There are several considerations when examining these results with respect to this limitation. First, freeway driving has a lower crash rate (Scanlon et al., 2024a). Including freeway driving makes this benchmark crash rate artificially lower, so, by including freeways in this study’s benchmark, the benchmark crash rate underestimates the true driving risk of where the Waymo ADS operates. Second, driving outside of these denser urban areas that the Waymo ADS operates would likely represent a reduction in overall relative crash risk. For example, commuters from the city would likely experience a reduced crash risk as they travel to less densely populated areas (Chen et al., 2024). Previous studies have shown that most injury collisions occur within a small radius from residency, and that American drivers rarely travel far from their place of residence, with approximately 80% of one-way household trips being less than 10 miles (DOE, 2022). Third, the benchmark drivers garaged in the Waymo deployment area are not operating with the same distribution of mileage within the geographical limits as the Waymo ADS. Chen et al. (2024) explored the effect of Waymo’s driving distribution on benchmark crash risk and found that - should the benchmark driving distribution match Waymo’s in San Francisco, Phoenix, and Los Angeles - the benchmark police-reported crash rates would have been between 14% and 38% higher. Due to all three of these limitations being expected to artificially suppress the benchmark crash rate (underestimation), the benchmarking results in this study are considered to be conservative. Surely, there is an opportunity in future work to leverage new data, such as insurance telematics, to more precisely define and leverage the benchmark driving exposure data to better account for this potential confounder.

Good to see studies of incidents in comparison to human drivers. Historically what I have seen is 'self-driving car crashes into X, which seems to be concentrating exclusively on outliers to prove a point.

Or even as a pedestrian. I'm sure AI will be fine to slow down if it sees a massive puddle and avoid soaking everyone in 3m radius just for lolz.