Security leads the whole lot we do at Waymo. This yr alone, Waymo has served over 700,000 ride-hailing journeys with public riders and no human driver. We couldn’t have hit that milestone with out placing security entrance and middle, and we’re working arduous to enhance the measurement, transparency, and efficiency of our fleet.

Our complete analysis — throughout greater than twenty security papers that we’ve revealed through the years to reinforce transparency — exhibits that the Waymo Driver performs safely throughout a variety of evaluations. Constructing on that work, we’ve revealed two new papers right now: one that compares the Waymo Driver’s crash charges to human drivers’ over our 7+ million rider-only miles from Phoenix, San Francisco, and Los Angeles; and one other that develops clear human crash benchmarks to allow such comparisons.

Our new analysis discovered that Waymo Driver efficiency led to a major discount within the charges of police-reported and injury-causing crashes in comparison with human drivers within the cities the place we function.

“These stories characterize a good-faith effort by Waymo to guage how the security of its autonomous driving system compares with the security of human driving. The outcomes are encouraging and characterize one step in our evolving understanding of autonomous driving security,” stated David Zuby, chief analysis officer of The Insurance coverage Institute for Freeway Security (IIHS) after reviewing the papers.

How the Waymo Driver compares to people

There’s no single metric that may present the security of autonomous driving. We use a number of totally different analysis metrics and strategies aiming to get a whole image of security. One in all many metrics that can be utilized to watch the rider-only service after it has launched is the variety of automobile crashes per mile of driving in comparison with people.

Within the efficiency research, we in contrast the Waymo Driver’s crash charges to human drivers’ on a number of totally different benchmarks. This research is likely one of the first to match general crash charges utilizing knowledge from absolutely autonomous operations solely, fairly than a mixture of absolutely autonomous driving and testing with a human behind the wheel. Not like the current analysis by Swiss Re that targeted on crashes leading to Waymo’s legal responsibility claims, this research contains all Waymo crashes, whatever the Waymo automobile’s position within the crash, and with any quantity of property injury. It additionally makes use of publicly obtainable knowledge, which permits different researchers to copy the outcomes.

Waymo’s knowledge was derived from crashes reported underneath NHTSA’s Standing Common Order (SGO), over 7.14 million absolutely autonomous miles pushed 24/7 by way of the tip of October 2023 throughout Phoenix, San Francisco, and Los Angeles. That knowledge was then in comparison with related human crash charges leading to police stories, accidents, and/or property injury.

When contemplating all areas collectively, in comparison with the human benchmarks, the Waymo Driver demonstrated:

• An 85% discount or 6.8 instances decrease crash fee involving any damage, from minor to extreme and deadly circumstances (0.41 incidence per million miles for the Waymo Driver vs 2.78 for the human benchmark)
• A 57% discount or 2.3 instances decrease police-reported crash fee (2.1 incidence per million miles for the Waymo Driver vs. 4.85 for the human benchmark)

Which means over the 7.1 million miles Waymo drove, there have been an estimated 17 fewer accidents and 20 fewer police-reported crashes in comparison with if human drivers with the benchmark crash fee would have pushed the identical distance within the areas we function.

When in comparison with crash charges in San Francisco, Phoenix, and Los Angeles, individually, the Waymo Driver considerably outperformed native respective human benchmarks as effectively (although the comparability in LA doesn’t but have sufficient mileage to be statistically vital). Notably, native human benchmarks different from one metropolis to a different — for instance, San Francisco had the best fee of crashes the place an damage was reported with 5.55 incidents per million miles, which was roughly thrice greater than the nationwide common.

P.c discount for the Waymo Driver in police-reported and any-injury-reported crashed automobile charges in comparison with human benchmarks

Generally, the Waymo Driver additionally demonstrated decrease property injury charges in comparison with the human benchmarks. Nonetheless, the benchmark charges themselves different significantly relying on the human knowledge supply, even inside the identical location, so warning ought to be taken when deciphering these outcomes.

Producing a legitimate comparability of human and AV efficiency

Figuring out legitimate and comparable human crash benchmarks is vital for understanding autonomous driving expertise’s efficiency. Regardless of public accessibility of each human and autonomous automobile (AV) crash knowledge, evaluating the 2 comes with its share of challenges. Our benchmarking paper goals to make sure a good comparability between AV and human driving by addressing the commonest errors and biases and establishing legitimate benchmarks from the cities by which Waymo operates.

There are two vital sources of statistical bias to manage for when evaluating human and autonomous driving. The primary is human crash underreporting. Whereas the information on human crashes that result in accidents or property injury is pretty sturdy, a lot of human low-severity crashes — like hitting some street particles or minor “fender benders” — usually are not reported to police. In distinction, AV corporations report even essentially the most minor crashes as a way to show the trustworthiness of autonomous driving on public roads. For instance, solely 21% of crashes that Waymo has reported to NHTSA thus far have resulted in a filed police report, whatever the social gathering at fault.

The second is variations in driving circumstances and/or automobile traits. Public human crash knowledge contains all street varieties, like freeways, the place the Waymo Driver at present solely operates with an autonomous specialist behind the wheel, in addition to varied automobile varieties from industrial heavy autos to passenger and bikes.

These variations imply that changes should be made to human crash knowledge earlier than evaluating it to AV crash charges. Simply as you wouldn’t examine biking in a road-race to mountain biking with out accounting for variations in altitude, terrain, and different elements, comparisons between human and autonomous driving have to account for variations too.
P.c discount for the Waymo Driver in police-reported and any-injury-reported crashed automobile charges in comparison with human benchmarks

To verify the human crash knowledge is legitimate and comparable, the damage and property injury benchmarks in our 7.1 million mile research both included underreporting changes for police-reported crashes or have been derived from naturalistic driving research databases, which equip autos with sensors to file driving and usually tend to catch low-severity crashes. The methodology we used was primarily based on finest practices from a literature evaluate of 12 previous research and 1 e book evaluating AV and human crash charges.

At Waymo, we imagine that worldwide standardization for legitimate evaluation of AV crash knowledge is required. We’re searching for alternatives to interact the neighborhood of visitors security consultants and requirements organizations to advance this work sooner or later.

Our complete strategy to security

Whereas being safer than human drivers in mixture is a crucial facet of AV security, it’s not enough by itself. The evaluation of crash charges is only one amongst many strategies used to guage the Waymo Driver’s efficiency, akin to how a routine well being checkup entails analyzing varied features of bodily well-being.

Our strategy goes past security metrics alone. Good driving habits issues, too — driving respectfully round different street customers in dependable, predictable methods, not inflicting pointless visitors or confusion. We’re working arduous to repeatedly enhance our driving habits throughout the board.

By way of these research, our aim is to offer the newest outcomes on our security efficiency to most people, improve transparency within the AV business, and allow the neighborhood of researchers, regulators, and lecturers learning AV security to advance the sphere.

SOURCE: Waymo