Waymo 安全影响
Waymo Safety Impact

原始链接: https://waymo.com/safety/impact/

评估像Waymo这样的自动驾驶系统(ADS)的安全性时,选择合适的比较指标至关重要。**车辆层级比率**——ADS车辆行驶里程中的事故数——是最合适的。将其与**事故层级比率**(所有车辆行驶里程中的事故数)进行比较可能会产生误导,因为单位不匹配。 例如,两辆车相撞的情景会产生不同的比率:0.5次事故/100英里 vs. 1辆受损车辆/100英里。错误地比较这些数据可能会错误地表明ADS的事故率更高。 同样,**人员层级比率**(行驶里程中的受伤人数)也存在问题。随着ADS车队的扩大,即使事故参与度保持一致,其人员层级比率也可能*下降*,仅仅因为更多的车辆贡献了总体里程。 由于这些偏差和数据限制(例如不完整的受伤报告),**车辆层级比率提供了最准确和可解释的比较**,用于衡量ADS性能与传统车辆安全基准。

## Waymo 安全性声明与讨论 Waymo 最近声称其车辆比人类驾驶员安全 13 倍,引发了 Hacker News 的讨论。 许多评论员认为 Waymo 的自动驾驶技术是一项重要的安全改进,但对数据的统计方法存在怀疑。 有人担心存在“选择性展示数据”的情况,以及缺乏在相同路线和条件下进行的比较。 尽管如此,一些用户分享了积极的经验,指出 Waymo 的注意力始终集中,反应速度比人类更快——从不分心或未能*发现*潜在的危险。 一位亚特兰大观察者报告说,即使在具有挑战性的道路上,驾驶也始终平稳。 有人提出了创新的用途,例如将 Waymo 车辆用作自行车骑手的“团队车”,提供安全和设备运输。 一些人指出,Waymo 目前的安全记录可能被其他司机因其车辆外观独特而格外小心所夸大,但最近的报告表明 Waymo 的驾驶行为有所改善,现在与熟练的人类驾驶员相似。
相关文章

原文

A crashed vehicle rate or vehicle-level rate is computed by counting the number of vehicles involved in crashes at a certain outcome level and dividing by the population-level VMT. For the Waymo crashes, the crashed vehicle rate is computed as the number of Waymo vehicles in crashes with a given outcome level divided by the total Rider-Only (RO) miles traveled by Waymo. For the benchmark, it is the total number of vehicles involved in crashes of a certain outcome in police report data divided by the total population VMT. 

Another metric available is a crash-level rate (i.e., number of crashes per population VMT). To illustrate why using a crash-level benchmark to compare to vehicle-level rate of an Automated Driving System (ADS) fleet creates a unit mismatch that could lead to incorrect conclusions, it’s useful to use a hypothetical, and simple, example. Consider a benchmark population that contains two vehicles that both drive 100 miles before crashing with each other (2 crashed vehicles, 1 crash, 200 population VMT). The crash-level rate is 0.5 crash per 100 miles (1 crash / 200 miles), while the vehicle-level rate is 1 crashed vehicle per 100 miles (2 crashed vehicles / 200 miles). This is akin to deriving benchmarks from police report crash data, where on average there are 1.8 vehicles involved in each crash and VMT data where VMT is estimated among all vehicles. Now consider a second ADS population that has 1 vehicle that also travels 100 miles before being involved in a crash with a vehicle that is not in the population. This situation is akin to how data is collected for ADS fleets. The total ADS fleet VMT is recorded, along with crashes involving an ADS vehicle. For the ADS fleet, the crashed vehicle (vehicle-level) rate is 1 crashed vehicle per 100 miles. If an analysis incorrectly compares the crash-level benchmark rate of 0.5 crashes per 100 miles to the ADS vehicle-level rate of 1 crashed vehicle per 100 miles, the conclusion would be that the ADS fleet crashes at a rate that is 2 times higher than the benchmark. The reality is that in this example, the ADS crash rate of 1 crashed vehicle per 100 miles is no different than the benchmark crashed vehicle rate, in which an individual driver of a vehicle was involved in 1 crash per 100 miles traveled.

This mistake of comparing a crash-level rate to a vehicle-level rate is easy to do when using aggregate statistics because summary statistics provided by research agencies often list the number of crashes instead of the number of vehicles involved in crashes. For example, Scanlon et al. (2024) reported that nationally there were 5,930,496 police-reported crashes in 2022, involving 10,528,849 crashed vehicles. The total national VMT for 2022 was  3.2 trillion miles. This means that the crash-level rate for the US is 1.9 crashes per million miles while the vehicle-level rate is 3.3 crashed vehicles per million miles. 

Another common metric used in traffic safety is injured people per VMT (i.e., a person-level rate). As a population level measure of the burden of crashes, a person-level rate has merit. There are several practical and interpretation issues that make a person-level rate not an ideal metric when comparing one population to another like is done in the Safety Impact Data Hub. A person-level rate for an ADS fleet operating in mixed traffic will appear to decrease as fleet size (or penetration) increases, even if crash involvement rate stays the same. Because crashes often involve multiple vehicles, the larger the fleet size the more likely it would be that multiple ADS vehicles are involved in a crash, which would decrease the person-level rate (same number of people involved in the crash, more VMT). This means that early in testing, the person-level rate of the ADS fleet would appear higher than the benchmark even if the ADS was involved in a similar number of crashes as the benchmark population. To address this bias, one could compute a fractional person-level rate defined as the total people involved in a crash at a given outcome divided by the number of vehicles in the crash. Although this fractional person-level rate addresses the bias in multiple vehicles, it creates a different bias in the interpretation of the results. The fraction person-level crash rate weights crashes involving fewer vehicles more than crashes that happen to involve multiple vehicles. There is also a practical limitation in that the NHTSA Standing General Order, the most comprehensive source of ADS crashes, reports only the maximum injury severity in the crash and not the number of injured occupants at given severity levels. So, it is not possible to compute a person-level rate from the SGO data today. This limitation also applies to some state crash databases, where only maximum severity is reported. Because of the potential biases in interpretation and reporting limitations, a vehicle-level rate is preferable to a person-level rate when comparing ADS and benchmark crash rates.

联系我们 contact @ memedata.com