Comparison of Waymo rider-only crash data to human benchmarks at 7.1 million miles
Level 4 - case-series / case-control
By design analogy, observational comparison of fleet operational data against historical/literature benchmarks.
OpenAlex W4403978380 · doi:10.1080/15389588.2024.2380786
What was done
The study evaluated the crash performance of the Waymo Driver, an SAE Level 4 automated driving system (ADS) operating in rider-only commercial ride-hailing fleets without human drivers present. ADS crash data were obtained from NHTSA Standing General Order (SGO) reports covering 7.14 million rider-only miles through October 2023 across Phoenix, Arizona; San Francisco, California; and Los Angeles, California. These ADS collision rates were compared against human driving benchmarks established in existing literature.
What was found
Across all locations combined, the any-injury-reported crash rate was 0.6 incidents per million miles (IPMM) for the ADS compared to 2.80 IPMM for the human benchmark, representing an 80% reduction (5 times lower crash rate; statistically significant overall and in San Francisco). Police-reported crash rates across all locations were 2.1 IPMM for the ADS versus 4.68 IPMM for the human benchmark, representing a 55% reduction (2.2 times lower crash rate; statistically significant overall and separately in San Francisco and Phoenix). For outcomes measuring any property damage or injury, 3 comparisons showed statistically significant decreases, while 3 were nonsignificant. Excluding ADS collisions with a delta-V under 1 mph removed approximately half of all ADS incidents, resulting in statistically significant reductions in all comparisons except one in San Francisco.
Why it matters
This paper provides empirical crash performance data for driverless commercial autonomous vehicles, demonstrating lower rates of injury and police-reported crashes compared to published human baseline rates in equivalent operating regions.
Limits
The study relies on external literature benchmarks for human driving rather than a directly concurrent, matched control cohort. Estimating human baseline rates for minor property damage carries high reporting uncertainty because human minor collisions frequently go unreported, confounding low-severity comparisons. Geographic exposure was limited to selected areas in three US cities.
Cited by
- partial Self-driving cars are currently about 10 times safer than human drivers.