There is this lingering question everyone keeps asking: are robotaxis actually safer? It is a bit like comparing apples to… well, very complicated, sensor-laden oranges. According to recent research featured on Misryoum, we have some fresh benchmarks to look at, specifically the data from Scanlon et al. (2024) and the newer Kusano et al. (2025) study. They’re using state police reports and vehicle miles traveled data from places like Phoenix, San Francisco, Los Angeles, and Austin to see if those autonomous cars are holding their own.
Here is the thing—you cannot just dump all the city stats into a pile and hope for the best. The researchers were careful to exclude freeways, focusing only on the road types where Waymo actually operates. Even the way they handle injury reporting is nuanced. They applied a 32% underreporting correction for minor injuries, based on that 2023 NHTSA study, while keeping the more serious incidents as raw, observed data. It sounds technical, but it’s basically just trying to account for the fact that humans aren’t great at reporting every minor fender-bender.
I remember walking past a Waymo parked in downtown SF, and I could swear I smelled ozone—or maybe it was just hot electronics. Anyway, it makes you realize how much tech is packed in there.
But wait, there is a catch. Not every street in a city is the same, right? If a robotaxi spends most of its time in the gnarly, high-traffic intersections while human drivers might stick to quieter residential lanes, that skews the crash numbers. The team used a method from Chen et al. (2024) to adjust for this. They call it spatial dynamic benchmarking. It essentially reweights the human data to better reflect the specific areas where Waymo puts in the most miles.
It is all part of what they call the RAVE—or Retrospective Automated Vehicle Evaluation—best practices. It’s an attempt to get the most honest comparison possible given the messiness of real-world data. Are the comparisons perfect? Probably not—no study really is. But it’s a lot better than just guessing.
Actually, I wonder if the city-level averages will ever be enough, or if we need to get even more granular. It’s an evolving process.