The Need for Connected Cooperative Autonomy in our Transportation System - Safety

A large part of my day-to-day traffic engineering work revolves around safety – vision zero plans, safe streets for all grants, intersection design, and safety studies.? Most of my academic research has revolved around autonomous vehicles, how to model them, their impact on roadway operations and how to weigh the decision to invest in them.? With all of the attention that AV safety received last year, I wanted to take a look at some of the data and incidents related to AVs and safety. While safety is one of the main long-term promises of AVs, what does the near-term look like?? I have seen a lot of articles and studies that are all pointing me to the same general story: We need to continue to steer our transportation system towards connected cooperative autonomy. Why and how we do this is important (and I will talk about that later), but the incidents and research that are coming out, when looked at from a broad perspective, paint a dangerous picture of agencies continuing to struggle with safety issues and increasing crash rates as long as humans are in control of motor vehicle operations. ?Connectivity and cooperation can increase the Operational Design Domains (ODDs) for autonomous vehicles as well as further increase the safety and efficiency of those operations.

How did I come to this conclusion?

Research shows that when humans are distracted or unengaged in the driving process, they are worse drivers and cause more crashes. We have seen a steady increase in the amount of distracted driving, but this may be at least partially a case of unintended consequences.? As the driving tasks continue to get easier, it requires less focus, making us more likely to try to multitask. A New York Times article discusses this as a possible explanation for the substantially higher rates of distracted driving in the U.S. compared to Europe (1). Drivers in the U.S. are much more likely to drive vehicles with automatic transmissions, which have a lower driver workload than the manual transmissions driven in Europe and allow for a free hand to hold a cell phone or perform other non-driving related tasks.? Driver-assist systems may have a similar effect; when you rarely have to brake or steer how much harder will it be to stay engaged in the driving process?? A recent study by LendingTree identified Tesla drivers as having the highest rate of accidents per driver (2).? This study does not normalize the crash rate by mileage driven or demographics and does not control for other variables, so Tesla ownership could be correlated to other high-risk factors for drivers. However, Tesla’s advanced driver assist systems might be making it easy for drivers to become complacent. We have seen numerous incidents that indicate drivers use these systems as if they are fully self-driving (rather than simply driver assistance) because of how often they are able to handle a situation. ?Subaru, which also has very advanced driver assistance systems was also in the top 3 highest crash rates per driver. The three brands with the lowest crash rates: Pontiac, Mercury, and Saturn, are all brands that were shut down before advanced driver assistance systems became prevalent. It is easy for drivers to be lulled into a false sense of security. If a system works 99% of the time, one might assume it will work 100% of the time and that the driver does not need to remain alert.? However, crashes and the situations that cause them are often rare situations, but situations we can only navigate successfully when we are at our most alert. As driver assistance systems continue to improve and become more prolific, it seems likely that people’s reliance on them will increase and society’s attention on the road will continue to diminish. ?

What about autonomous vehicles, how are they doing on safety?

Since National Highway Traffic Safety Administration NHTSA issued the standing General Order on Automated Driving Systems (ADS) that requires crash reporting, we have a consistent set of crash data to evaluate (3). With the increased deployments of autonomous vehicles, we have more miles driven. From a data perspective, this means that we have a larger sample size of both miles driven and crashes for use in analyses, making the results more reliable.? There have been at least two large studies of AV safety conducted in recent years. The most comprehensive study to date was self-performed by Waymo based on over 7.1 million miles (4).? Cruise also self-reported crash rates at 1 million miles driven (5). These both showed decreases in rates of injury crashes compared to manual-driven vehicles. ?While these studies do not appear to be published in peer-reviewed journals yet, the methodology is sound. ?The researchers are very transparent in the study limitations, including that available data is not sufficient to control for route choice, driver characteristics, variances across times of service, etc. Still, these present additional data points in the overall picture that indicate that AVs can, and currently are, achieving safer outcomes when they are thoughtfully deployed in the Operational Design Domains (ODDs) that they are equipped to handle.? The comparisons were conducted based on different human benchmarks for safety performance. While multiple organizations (Waymo (6) and University of Michigan Transportation Research Institute (UMTRI) (UMTRI) (7)) have developed benchmarks for human driving based on different data sets and crash types, these studies showed significant safety benefits compared to all benchmarks. The statistical significance of these diminishes as we move from all crashes to only injury crashes, as the incidents per million miles traveled significantly reduce.

But we have heard a lot about AV crashes…?

There have been some highly-publicized incidents that could lead to the perception that AV technology may decrease, rather than increase, safety. ?Last year, a Cruise vehicle injured a pedestrian by dragging her under the vehicle. This was a tragic incident, but one that I think speaks to the need for autonomous vehicles. Many media reports suggested a human driver would never have done this, yet, it’s easy to find media reports of human drivers striking and dragging pedestrians with fatal outcomes. (The incident with the Cruise vehicle did not result in a fatality). ?The chain of events in the Cruise crash was initiated when a human-driven vehicle struck the pedestrian, knocking them into an unavoidable collision with the Cruise vehicle. ?Had that first vehicle been an AV, it is very likely the entire collision and all injuries would have been avoided. ?It was absolutely a shortcoming of the autonomous driving system that it failed to track the pedestrian correctly after contact, recognize the pedestrian had been trapped under the vehicle, and remain stationary rather than engaging in the standard post-crash behavior to pull off of the road. While this outcome is unacceptable, one of the promises of AVs is that they can and will get better as they learn from their experience. ?Cruise can determine the cause of this failure and improve their driving system to avoid this in the future, making their self-driving algorithm safer moving forward. Other AV companies will also take lessons from this experience and improve their sensors and algorithms. This is different from human driving incidents, where one driver’s error does not lead to better or safer driving across all humans. ?

While recent data indicates that autonomous-driven vehicles are safer on average than human drivers, they are not perfect. The hard reality is that AVs will likely be involved in different crashes than humans, and while that doesn’t make it any better for the person who was injured in the AV crash, we also have to consider how many lives are saved by crashes avoided by AVs and how many friends and families don’t have to experience that grief and loss. What if we change the paradigm and look at every human-driven crash and ask ourselves if an autonomous vehicle would have avoided that crash with the same scrutiny that we look at AV crashes that a human might have avoided? How much more of a push would there be to implement the infrastructure and the policies to advance and improve autonomous driving? ?A similar analogy might be the warning sign that is put up on a curve.? That sign helps many people avoid a crash, but if a vehicle goes off the road and hits that sign, they may be more seriously injured than if that sign was not present. However, if that sign prevented more injuries than it caused, we consider it to be a positive safety improvement.

I feel that the concern for this incident should be more about the transparency around the crash than the crash itself. We need to understand and trust these AVs and the companies that provide them.? When an incident happens, the data need to be fully shared in a timely manner with the regulating agencies. According to the findings of a law firm hired by Cruise to review how the firm handled this matter, “While Cruise employees clearly demonstrated mistakes of judgment and failure to appreciate the importance of transparency and accountability, based on Quinn Emanuel’s review to date, the evidence does not establish that Cruise employees sought to intentionally mislead government regulators” and “Cruise’s senior leadership repeatedly failed to understand the importance of public trust and accountability.”? I hope that this again is an area where Cruise and all other AV companies will learn from this so we can continue to build public trust around this life-saving technology.?

Check back for future articles on the need for connected cooperative autonomy in our transportation systems, and please reach out if you would like to discuss AVs or traffic safety.

1)???? Why Are So Many American Pedestrians Dying at Night? https://www.nytimes.com/interactive/2023/12/11/upshot/nighttime-deaths.html

2)???? LendingTree study https://www.lendingtree.com/insurance/brand-incidents-study/

3)???? Standing General Order on Crash Reporting https://www.nhtsa.gov/laws-regulations/standing-general-order-crash-reporting#ads

4)???? Comparison of Waymo Rider-Only Crash Data to Human Benchmarks at 7.1 Million Miles https://assets.ctfassets.net/e6t5diu0txbw/54ngcIlGK4EZnUapYvAjyf/7a5b30a670350cc1d85c9d07ca282b0c/Comparison_of_Waymo_Rider_Only_Crash_Data_to_Human_Benchmarks_at_7_1_Million_Miles_arxiv.pdf

5)???? Cruise’s Safety Record Over 1 Million Driverless Miles https://getcruise.com/news/blog/2023/cruises-safety-record-over-one-million-driverless-miles/

6)???? Benchmarks for Retrospective Automated Driving System Crash Rate Analysis Using Police-Reported Crash Data https://arxiv.org/pdf/2312.13228.pdf

7)???? Establishing a Crash Rate Benchmark Using Large-Scale Naturalistic Human Ridehail Data https://deepblue.lib.umich.edu/bitstream/handle/2027.42/178179/UMTRI-2023-18.pdf?sequence=4&isAllowed=y

8)???? Report To The Boards of Directors of Cruise LLC, GM Cruise Holdings LLC, and General Motors Holdings LLC Regarding The October 2, 2023 Accident In San Francisco https://assets.ctfassets.net/95kuvdv8zn1v/1mb55pLYkkXVn0nXxEXz7w/9fb0e4938a89dc5cc09bf39e86ce5b9c/2024.01.24_Quinn_Emanuel_Report_re_Cruise.pdf