.
Death By Uber 1: When Self-Driving Car Kill
Marc Green
In 1896 Bridget Driscoll became the first pedestrian to die in a car collision.1 She was visiting the grounds of the Crystal Palace in London, where the Anglo-French Motorcar Corporation was demonstrating its new machine. Driver Arthur Edsel said that before striking her he had shouted and had rung his bell furiously in a futile attempt to warn Ms. Driscoll, who just froze and seemed "bewildered." Despite the anger and courtroom histrionics of Ms. Driscoll's family, the coroner ruled the to be death accidental and said that he hoped Ms. Driscoll's tragedy would be the last of its kind. It wasn't.
In 2018 the conflict between vehicles and pedestrians entered a new phase when Elaine Marie Herzberg became the first pedestrian killed
in a collision with a self-driving vehicle. She will doubtless not be the last. The increasing number of "self-driving" cars will
likely make such collisions commonplace. When such a collision occurs, the question arises of who is at fault. If the car company
is found liable, the entire industry may fail under a mountain of law suits. It is not an exaggeration to say that the fate of the
entire industry depends on how cases such as Ms. Herzberg's death are adjudicated. If the car company is held liable, then it may be that self-driving cars will have to be perfect, or they simply won't be. The stakes are high.
Uber owned the car that struck Ms. Herzberg, but the state of Arizona assigned no blame to Uber and allowed them to escape any liability, a decision some viewed as based on economics rather than on public safety (e.g., Smiley, 2022).
Instead, the local authorities, the Tempe Police Department (TPD), and the National Transportation Safety Board (NTSB) issued
reports saying that the vehicle "driver," Rafaela Vasquez, caused the collision.
The public release of a video showing Ms. Vasquez looking away from the road just prior to collision unleashed a
torrent of scathing condemnation on social media sites.2 This July, Ms. Vasquez agreed to plead guilty to charge of "endangerment."
There have been many articles about Ms. Herberg's death, written almost exclusively by journalists. Here, I provide the
first in-depth human factors review and examine the TPD and NTSB reports as well as the collision itself.
Both the TPD and the NTSB reports are severely flawed but in different ways. The TPD performed a detailed accident reconstruction,
but conducted misguided tests and blindly performed cookbook analyses that the report author didn't really understand.
Worst of all, the report completely ignored the large documented performance losses of supervisors3
of self-driving vehicles compared to drivers of conventional vehicles. The NTSB report performed a less detailed,
but largely competent analysis that emphasized the performance losses experienced by self-driving car supervisors.
Yet the NTSB report somehow reached a final conclusion that directly contradicted their own analysis and agreed with the TPD that Ms. Vasquez caused the crash.
The analysis is divided into four main parts.
- This part reviews the basic facts of the collision derived primarily from the TPD and NTSB reports,
although Lauren Smiley's 2022 Wired article provided helpful context;
- The second reviews the TPD report, which outlines their detailed accident reconstruction.
The TPD's goal was to determine if the collision were avoidable so that blame might be established.
The review explains why the TPD's determination of both sight distance and perception-reaction time (PRT) are badly flawed for many reasons.
The discussion also mentions some aspects of the NTSB report where they depart from the TPD.
The discussion highlights the many pitfalls of using research data to analyze real-world situations;
-
The third part continues the critique of the TPD report by focusing on a more general fundamental flaw-the accident reconstruction was based on data and analyses for conventional drivers.
The discussion describes the evidence, some discussed by the NTSB, that self-driving vehicles create "automation complacency" which results in decreased the vigilance and increase response times in supervisors such as Ms. Vasquez.
This leads to the inescapable conclusion that self-driving vehicles put highly "impaired" persons in the "driver's seat" should the automation fail; and
-
The final part focuses on the NTSB report which sought to determine causation.
It concluded that although the collision had many contributing factors, Ms. Vasquez was the "probable cause."
The discussion explains why this conclusion contradicts their own analysis and why safety organizations such as NTSB and NHTSA, as well as laypeople, are heavily prone to attribute causation for mishaps to humans at the "sharp end" of the system.
The explanation lies in the nature of causal attribution and in a range of cognitive biases. It explains why the leak of video of Ms. Vasquez looking at the road was highly irresponsible because it was designed to make her the culprit
and to remove the spotlight from those really responsible. Lastly, I list some conclusions
about the reports and the collision itself as well some general points raised by this collision about the dangers posed by self-driving vehicles and the scapegoating of "drivers" when the automation fails.
Facts of the Collision
The collision occurred in Tempe, Arizona on March 18, 2018. At about 9:46 PM, Rafaela Vasquez was supervising a 2017 Volvo XC90 TB equipped with Level 3 automation.4
This is a "conditional" automation level where the car drives itself although it might make takeover requests to the supervisor. In the parlance of the self-driving literature, the vehicle had an "automatic driving system (ADS). Such systems can only operate over a limited range of conditions termed the "operational design domain" (ODD). When the ADS reaches its ODD, it is supposed to issue a "takeover request." (TOR) to the supervisor. The supervisor's time interval to respond is the "takeover time" (TOT).
She was traveling in the northbound right lane of North Mill Avenue towards the intersection of Washington Street/Curry Road. Figure 1 shows that the roadway had a shallow curve to the left and passed through an overpass. This section of Mill Avenue was part of a designated circuit that Ms. Vasquez was traversing for the 73rd time and the third circuit of the day.
At about the same time Elaine Marie Herzberg was walking her bicycle eastbound toward North Mill Avenue about 400 feet south of the intersection while under the influence of drugs. Her clothing was a dark top and what appears in videos to be blue jeans.
She was walking in front of a bicycle, which was red/black with no lights or side reflectors. Red is a low visibility color in dim light, so the bicycle frame was not conspicuous.5 When entering the right northbound lane #2, the Uber Volvo and Ms. Herzberg collided, resulting in her suffering fatal injuries.
Figure 1 shows an overview of the roadway leading to the point of impact. The road was one-way with traffic flowing northbound. It was level and was slightly curved leftward until reaching an overpass where it began to straighten out. The collision occurred 243 feet into the straight segment according to the NTSB report. At the point of collision, the road had for lanes, two left-turn lanes and two through lanes. The road's posted speed was 45 mph, and the police report coded the road illumination as "dark, lighted."
Figure 1 The roadway leading up to the point of impact. The red arrow points to Mill Avenue which was one-way north.
The vehicle sensors detected Ms. Herzberg 5.6 seconds prior to impact. Figure 2 (taken from the National Transportation Safety Board report) shows that she was then located just slightly left of the boundary between the two left-turn lanes. During the 5.6 seconds between detection and impact, the vehicle's software was unable to conclusively identify the detected object as a pedestrian (or bicycle) and did not alter vehicle speed or trajectory. About 1.2 seconds prior to impact, when Ms. Herzberg was already in lane #2, the vehicle finally concluded that the object ahead required avoidance. At this point, it still did not alert Ms. Vasquez or change speed or trajectory due to a software protocol that prevented braking beyond a set criterion (7.1 g) for one second. (This was a feature designed to prevent false alarms and unnecessary braking.) Finally, the vehicle signaled a takeover request (TOR) at a time-to-collision (TTC) of 200 msec (1/5 second) prior to collision, which made the takeover time (TOT) impossibly short. The impact occurred 8.14 feet into the travel lane, according to the TPD report. Since the lanes are about 13.65 feet wide (according to the NTSB), Ms. Herzberg was about 2*13.65 + 8.14 = 35.44 feet from the point of impact when the Uber car detected her.6 She covered that distance in 5.6 seconds so her average speed was 35.44/5.6=6.32 ft/sec. This is within the high end of normal walking speed.
Figure 2 The location of Ms. Herzberg in the seconds leading up to the collisions. From Figure 1 in the NTSB Final report.
Videos and eye movement recordings show the Volvo's supervisor Rafaela Vasquez looking downward and rightward during much of the time leading up to collision and only responding at a fraction of a second before impact. The vehicle had a touch pad to log anomalous events at roughly the fixated location but there is no indication that she was interacting with it. There were also cell phones located in her general direction of gaze. One phone was showing a video on the streaming service Hulu. However, Ms. Vasquez stated that she was not watching the phone but only listening to it. Instead, she was monitoring a different phone that had the Uber Advanced Technologies Group's (UATG) Slack app, which was used to communicate with managers at the base. She was performing a task required by the test protocol.
Ms. Vasquez had no "co-pilot." She was alone in the vehicle at the time. Uber had previously run their test vehicles with two occupants. The person in the left (driver) seat would call out potential obstacles, such as bicycles, and street signs while the person in the right (passenger) seat viewed a laptop that confirmed system detection. In a move likely aimed at saving money, UATG removed the passenger and a single person was left to supervise the vehicle and buttons on the touch pad to indicate log-worthy events.
Some at the USTG were concerned that the monotony and boredom of a lone driver would produce fatigue and lead to supervisors to become easily distracted. They proposed that the video camera facing the supervisor could be randomly monitored in real-time or later to check whether their eyes stayed on the road. However, such checks were not performed. At least one USTG supervisor urged the return of the passenger after finding that vehicles with single occupants often came back from tests with damage (Smiley, 2022). Ms. Herzberg's collision occurred only a few days later.
Endnotes
1There was a previous death from an automobile-human collision. In 1869, Mary Ward of Ireland was a passenger in a car that hit a bump. She fell out and was run over. However, she was not a pedestrian in any conventional sense.
2The video is available for online viewing at https://yewtu.be/watch?v=cRSMYlFqUEA.
3The NTSB report variously describes Ms. Vasquez as the "driver" or the "operator."
In fact, she was neither since she was neither driving nor operating anything in conventional meaning of the terms. At level 3 automation, the car is presumed to drive itself with a human only intervening when necessary, usually in response to a takeover request (TOR) that occurs when the automatic driving system (ADS) reaches the limit of its "operational design domain" (ODD).
The SAE even says that at level 3 and above "You are not driving when these automated features are engaged."
There is no universally accepted substitute for "driver" which has been applied to the person monitoring a self-driving car,
but I will use the more accurate term "supervisor" which is becoming increasingly common in the automated driving literature.
However, I sometimes refer to the supervisor as the "driver" in quotes when making an ironic comparison.
4The automation was designed to operate autonomously on a pre-specified eleven mile loop that the Uber Volvo
traveled repeatedly. Most details of the technology for accomplishing the task are beyond the scope of this article and are not discussed. However, a few aspects are relevant.
5Rather than continually saying "pedestrian and bicycle," I will simply refer to the "pedestrian," but it
should be understood that I mean both.
6Ms. Herzberg's path was not quite a straight right angle to the roadway. However, calculations show that the difference from 90 degrees and linearity was so small that their effects on walking speed estimate are negligible.
Next: Death By Uber Part 2: The Police Report
A division of 2057949 Ontario, Inc.
Copyright © 2024 Marc Green, Ph. D.
Home Page
Contact Us
A division of 2057949 Ontario, Inc.
Copyright © 2024 Marc Green, Ph. D.
Home Page
Contact Us