Descriptive analysis of reports on autonomous vehicle collisions in California: January 2021–June 2022

: The characteristics of autonomous vehicles’ collisions from 2021 and the first half of 2022 in California confirm trends reported in previous years. Driving in autonomous mode was associated with fewer instances in which the AV was deemed to be at fault in a collision. Most collisions in autonomous mode were rear-end collisions at intersections. Single vehicle collisions occurred mostly in manual mode. Collisions with vulnerable road users occurred mostly while the autonomous vehicle was in manual mode, often right after disengagement from autonomous mode. In collisions with other vehicles that occurred after disengagement, the other vehicles were frequently deemed to be at fault. Compared to 2021, the collision reports from the first half of 2022 indicate higher shares of collisions in autonomous mode, rear-end collisions, and collisions with vulnerable road users.


Introduction
A part of the EU's executive plans is to propose by the end of 2022 the world's first technical legislation that will allow member states to approve the registration and sales of limited numbers of vehicles with advanced, SAE level 4, self-driving technology (Posaner 2022). As safety is a major concern in deploying such autonomous vehicles (AVs) into complex traffic environment, knowledge and learning from reported AVs collisions are essential.
Currently the most comprehensive data on AVs collisions have been gathered by the Department of Motor Vehicles (DMV) in California, US. Since 2014, DMV has administered the Autonomous Vehicles Program and issues permits to companies that test and deploy AVs on California public roads. AVs tested in California are typically passenger cars, mostly operating under 'conditional driving automation', i.e. SAE level 3 and recently level 4 (SAE 2021). The program requires all collisions ('any collision that resulted in property damage, bodily injury, or death within 10 days of the incident') and disengagements ('disengagements from autonomous mode because of technology failure or situations requiring the test driver/operator to take manual control of the vehicle to operate safely') to be reported to the DMV. As of August 23, 2022, the DMV has received 506 Autonomous Vehicle Collision Reports (DMV 2022). The 2 collected data are publicly available from an online database. This database has frequently been harvested by researchers to explore various aspects of AVs' collisions. Results from previous studies show consistently that AV collisions typically are of relatively low severity and are often rear-end collisions (see Table 1 for an overview).  Dixit et al. (2016Dixit et al. ( ) 12 collisions 9/2014Dixit et al. ( -11/2015 No serious injuries reported. All collisions were either rear-end or side-swipe. System failure was the most common cause of disengagement. Favarò et al. (2017) 26 collisions 9/2014-3/2017 Rear-end collisions were the most frequent collisions type. In 85% of collisions, AV was not-at-fault. Biever et al. (2019) 115 collisions 9/2014-6/2019 Rear-end collisions (AV being struck) most frequent (59%). 30% of collisions were side impacts (AV being struck). 9% of collisions related to disengagement. 92% of collisions in very low speed, 14% of collisions resulted in an injury (mostly minor). In collisions with AV in automated mode, non-AV's is more likely to be at-fault. Rear-end collisions (AV being struck) most frequent, followed by being sideswiped. More severe injuries on the highway. Rear-end and sideswipe collisions were predominant collision types. Stopped AV and conventional car proceeding straight is the common pre-collision scenario. The severity of AV collisions is lower than in collisions with conventional cars. Ye et al. (2020) 133 collisions 1/2017-6/2019 71% of the injured were AV occupants (head and neck were the most commonly injured). Driving in poor lighting was associated with greater injury risk. Collisions with VRU or incidents happening during commute periods led to more injuries. Human-initiated disengagements are more likely than ADS initiated disengagements. ADS-initiated disengagements are mostly due to planning and hardware/software discrepancies and occur on streets and roads environments than on high-speed facilities. Boggs et al. (2020b) 124 collisions 10/2014-11/2018 61% of collisions were rear-end collisions, with higher likelihood in mixed land-use settings. 13% of collisions were injury collisions. 10% of collisions were related to disengagement. Injury collisions are less likely at roadways with marked centrelines and in clear weather. Das et al. (2020) 151 collisions 9/2014-5/2019 Current narrative documentation is not sufficient in determining the driving mode. Classes associated with turning, multi-vehicle collisions, dark lighting conditions with streetlights, and sideswipe and rear-end collisions were associated with a higher proportion of injury severity level.  Das et al. (2020) 151 collisions 9/2014-5/2019 Current narrative documentation is not sufficient in determining the driving mode. Classes associated with turning, multi-vehicle collisions, dark lighting conditions with streetlights, and sideswipe and rear-end collisions were associated with a higher proportion of injury severity level. Petrović et al. (2020) 53 collisions 2015-2017 64% of AV collisions were rear-end collisions (vs. 28% for conventional car crashes). Collision types 'broadside' and 'pedestrian' represented a total of 5.7% of AV collisions, while 42% of conventional car collisions. Goodall (2021) 256 crashes 10/2014-3/2020 AV in autonomous mode were struck from behind at 4.8 times the rate of conventional car in a naturalistic driving study. AVs were more likely to be struck when stopped than when moving compared to conventional cars. Bicyclists and scooterists are more likely to be involved in the AV collisions directly, and bicyclists are likely to be at fault. Pedestrians appear more in the indirectly involvements. collisions that involve VRUs indirectly are likely to occur when the AVs are in autonomous mode. Crosswalks, intersections, traffic signals, movements of AVs are the key predictors of the VRUs-AV related collisions. * This study included also six AV collisions outside California that were reported by the National Transportation Safety Board (NTSB). In the table we include only the findings relevant to collisions from DMV database.

Leilabadi & Schmidt
Because of the rapid development in AV technology and other factors such as increased walking and cycling in urban areas, new patterns might emerge in AV collisions. Therefore, in our study we utilized the DMV database to explore whether recent collision data (i.e. collisions reported in 2021 and in the first half of 2022) provide any evidence for changes. Additionally, we compared characteristics of collisions where the AV was in manual mode, autonomous mode, or right after disengagement.

Descriptives
This study utilizes DMV collisions reports from period January 2021 to June 2022. Using descriptive statistics, it explores the following variables from the DMV database: • Collision type and location • Type of road user(s) involved • Severity levels • AV mode in collision (manual, autonomous, after disengagement) • Road user at-fault (estimated from collision narratives) • Million vehicle•kilometres travelled (MVKT)-available only for 2021.

Odds ratios
For comparing shares of collision properties between types of collisions we calculated odds ratios with 95% confidence intervals according to Christensen (1990).

Time periods
In 2021, eight companies reported 117 collisions to the DMV. The total distance driven by these eight companies was 6.06 MVKT (0.02-3.74 range). These were the highest annual values reported to the DMV database so far, both for collisions and distance travelled. In the first half of 2022, the testing companies reported 91 collisions to DMV. No data on MVKT is available for 2022.

Mean collision rate
The mean collision rate of AVs in 2021 was 19.3 collisions per MVKT, which is very similar to earlier studies. The smaller companies (1-9 reported collisions) had a higher mean collision rate (24.6 per MVKT) than the two largest companies that together account for 85% of MVKT (30 and 64 collisions, 18.2 collisions per MVKT).   Figure 1 shows collision types and other collision characteristics by AV mode for the whole studied period, while Table 2 compares selected characteristics of AV collisions between 2021 and the first half of 2022.

Primarily autonomous mode
In most collisions, the AV was in autonomous mode (93 collisions, 45%), followed by manual mode (86 collisions, 41%). The remaining 29 collisions (14%) occurred right after a disengagement. The share of collisions in autonomous mode increased from 38% to 53% from 2021 to 2022, the share of disengagement collisions decreased from 19% to 8%, while the share in manual mode was about unchanged (decreased from 43% to 40%).

Odds for being at-fault
When in autonomous mode, the odds of being at-fault was reduced by 78% compared to other modes (95%, CI 1 [-91%; -46%]). Since being at-fault was assessed based on AV-company reports, there may be some bias in the assigned responsibility in favour of AVs.

Severity levels
Most collisions involved no injury (88%) and only minor or no property damage (85%), without a substantial difference between modes and years. From 24 injury collisions, 22 (92%) involved minor injuries; the remaining two collisions resulted in a moderate injury.
Most (n = 16; 67%) injury collisions involved a car. AV was in an autonomous mode in 56% of these collisions. Most injury collisions with cars were rear-end collisions on intersections when the AV stopped prior to the collision and then was hit by another car.
VRU (mostly cyclists) were involved in eight injury collisions (33%). AV was in motion in all these collisions, and in all but one AV was after a disengagement or in automated mode. There was no obvious prevailing type of collision with VRU.

Collision locations
Most collisions occurred at intersections, both in 2021 and 2022. In autonomous mode the odds of a collision occurring at an intersection are far higher than of occurring at other locations (124%, CI [26%; 300%]). Figure 2 shows that most collisions were rear-end collisions, followed by sideswipes. When all modes are regarded together, the share of rear-end collisions increased from 40% in 2021 to 52% in 2022 (45% in total). However, the share of rear end collisions in autonomous mode was about unchanged (62% in 2021 vs. 69% in 2022).

Collision types
Rear-end collisions in autonomous mode are occurred typically at intersections where another vehicle was rear-ending the AV. Such collisions occur often because the AV brakes abruptly as the traffic signal switches to yellow, while following drivers expect it to continue until the light switches to red (Biever et al. 2019). In autonomous mode the odds of a collision being a rear-end collision is far higher than otherwise (374%, CI [163%; 753%]).
Sideswipes occurred equally often at intersections and mid-blocks. Besides single vehicle collisions, sideswipes are also the collision type where AVs were most often at-fault (20% of all sideswipes vs. 6% of rear-end collision). Remarkably, in most sideswipe collisions in which the AV was deemed to be at-fault (six out of nine collisions), the AV was in manual mode and hit a parked vehicle. In most single vehicle collisions, the AV was in manual mode (57%). This indicates that automated mode is most reliable in the absence of other road users. All single collisions were with fixed objects.

Collisions with VRU
In collisions between an AV and a vulnerable road user (n = 20), the VRU (cyclist, skateboarder or e-scooterist) was at-fault in 16 collisions (80%); for the remaining four VRU-collisions information about who was at-fault is missing.
Most VRU-collisions occurred at intersections (75%). From 2021 to 2022, the share of VRUcollisions is about doubled (from 7% to 13% of all collisions, based on 8 and 13 VRU collisions in 2021 and 2022, respectively).

Collisions in disengagement mode
After disengagement, the odds of being at-fault was about as high as in manual mode (-31%, CI [-75%; +92%]). Typical disengagement collision types were sideswipe and angle midblock collisions where the AV driver disengaged to solve a critical traffic situation. VRUs were overrepresented in disengagement collisions.

Conclusions
Our findings confirm results from earlier studies regarding the most frequent collision types and consequences. Additionally, they show that driving in autonomous mode was significantly related to rear-end collisions and other collisions where the collision partner is at-fault, but not to single collisions. A possible explanation is that AVs in autonomous mode 'behave' differently from manually driven cars, and thus more unpredictable for other road users.
The total number of collisions per six months is about doubled from 2021 to 2022, which may be attributed to increased AV testing. Collision rates could not be compared due to missing data on MVKT from 2022.
A comparison between 2021 and 2022 indicates that the share of collisions in autonomous mode and the share of rear-end collisions have increased, although there was no increase in the share 7 of rear-end collisions in autonomous mode. The increase of collisions in autonomous mode may be related to the development of AV technology, which allows driving in automated mode more frequently.
The share of VRUs involved in collisions has also increased from 2021 to 2022. Although the comparison is based on relatively few collisions, this requires extra attention, as AVs are going to encounter VRUs more and more frequently. VRUs present specific challenges for AVs, because they are often less predictable and more difficult to detect than cars.

Declaration of competing interests
The authors report no competing interests.