Could Disengagement Reports Indicate Evolution of Autonomous Vehicles?
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. There is a missing space in the text for 'Madetremendous progress';
2. It is recommended to label the units of the numbers in Figure 1;
3. It is recommended to adjust the width of certain columns in some tables to make them more aesthetically pleasing and increase readability;
4. The format of the table lacks uniformity;
5. It is suggested to increase the number of references and adjust the format to be consistent;
6. Please add the author's viewpoint and explain it in detail in the summary section of the article.
Author Response
Comment 1: There is a missing space in the text for 'Madetremendous progress';
Answer 1: The missing space were adedd, also few others mistakes we corrected.
Comment 2: It is recommended to label the units of the numbers in Figure 1;
Answer 2: The description of the units (individuals) was cut to the description of the Y-axis.
Comment 3: It is recommended to adjust the width of certain columns in some tables to make them more aesthetically pleasing and increase readability;
Answer 3: Tables were adjusted so their readibility and aestheticality is increased.
Comment 4: The format of the table lacks uniformity;
Answer 4: Uniformity to the tables were added.
Comment 5: It is suggested to increase the number of references and adjust the format to be consistent;
Answer 5: The references were supplemented by another 35 for a final number of 50 references.
Comment 6: Please add the author's viewpoint and explain it in detail in the summary section of the article.
Answer 6: The view of the authors' team is reflected in the Conclusion, which in the new version contains an interpretation of the results, further possible context, specific recommendations including a proposal for further research.
Reviewer 2 Report
Comments and Suggestions for AuthorsIntroduction
The introduction of this manuscript lacks a scientific structure. Few citations are provided, in several parts. From line 39 to 60 there are no significative or remarkable topics. It looks like a magazine or newspaper report of events.
L. 64-64 Is the prediction by Garza or LV? Is it possible that LV is the name of the author?
Cerny et al., 2022 should be the adequate citation, authors are more than two.
What kind of autonomous vehicles are you addressing by saying that currently there are thousands. Please explain this. Where are they deployed? Are you talking about fully autonomous? This is an important point, but it seems incomplete and voluntary confusing. You must define uniquely what the Autonomous vehicles are for your manuscript. Be precise on this point.
Improve the quality of Figures. Figure 1 does not have the same font of the text, has all the years merged. You have enough space to improve the quality of this figure.
L. 85 – 107 are still dealing with autonomous vehicles (not solved the ambiguity about the level of automation of the vehicles) and regulations. You have to report the precise details of the regulations mentioned to let readers use them. This part jumps out and it is not well-linked to the rest.
Table 1 is not self-explaining. You have to interpret the table while reading.
Figure 3 can substitute Table 1. In accordance with this figure, it is possible to state the most frequent boundary conditions occurring when the system disengaged.
Analysis of Autonomous Vehicles Disengagements
You cannot deal with autonomous vehicles generally when assessing specific disengagements. You have to create clusters from L2 vehicles to L5 vehicles. After this clustering process, you can create manufacturer clusters according to each type of vehicle.
Figure 6 is not useful. You could use kilometers and disengagements, not cumulating them. Currently it is not explaining and the regression line is useless. Takeovers or disengagements?
Same corrections for figure 7. Is it possible that the two figures show the same distribution having different variables? This means that the variables are correlated. More statistical analysis should be done.
Figures 8-10 can be useful but as they are presented, they do not contribute to the improvement of manuscript quality.
What does the safety driver disengagement mean? The need for driver to be more reactive than the system, the fear of the driver that the system does not perform well its task and then react before anything can be done by the system? There are several flaws in this discussion.
You mentioned different types of disengagements, why are there only two in table 2?
Bullet point list below the table is not adequate without any line to introduce.
Can you divide the results according to the level of automation?
You can use up to the second degree of polynomial regression to be consistent. Greater rates are undoubtedly forced. Statistically and mathematically, how can you represent a fourth polynomial curve? Which are the most recurrent events and phenomena that you can reproduce by a sixth-fifth grade polynomial curve?
L. 340 - 355 Why did not investigate these other factors? Curves are similar, you have to choose the most representative ones. Including all the curves make the figure not significantly different and not remarkable. Moreover, very poor mathematical and statistical approaches are used, with very low correlations that do not explain the variability of the phenomenon.
Discussions
Discussing results should mean that you have to assess the novelties of your results, comparing them with existing ones and improving the discussion with meaningful insights. This is a short report of the results.
Conclusions
This does not conclude properly.
References
Citing just 15 fonts for one of the most discussed topic in the field of transportation safety is not adequate.
Author Response
Comment 1: The introduction of this manuscript lacks a scientific structure. Few citations are provided, in several parts. From line 39 to 60 there are no significative or remarkable topics. It looks like a magazine or newspaper report of events.
Response: 1: 12 new sources added (13 citations)
Comment 2: L. 64-64 Is the prediction by Garza or LV? Is it possible that LV is the name of the author?
Response 2: The prediction is offered in Garza, 2011. Corrected in the manuscript. Yes, the author's full name is Chen Lv.
Comment 3: Cerny et al., 2022 should be the adequate citation, authors are more than two.
Response 3: Corrected citation across the manuscript.
Comment 4: What kind of autonomous vehicles are you addressing by saying that currently there are thousands. Please explain this. Where are they deployed? Are you talking about fully autonomous? This is an important point, but it seems incomplete and voluntary confusing. You must define uniquely what the Autonomous vehicles are for your manuscript. Be precise on this point.
Response: 4: A focus on the specific level of autonomous vehicles and further description is added in the Introduction section.
Comment 5: Improve the quality of Figures. Figure 1 does not have the same font of the text, has all the years merged. You have enough space to improve the quality of this figure.
Response: 5: Mentioned figure and few others quality and uniformity were improved.
Comment 6: L. 85 – 107 are still dealing with autonomous vehicles (not solved the ambiguity about the level of automation of the vehicles) and regulations. You have to report the precise details of the regulations mentioned to let readers use them. This part jumps out and it is not well-linked to the rest.
Response: 6: A more detailed text has been added in the relevant text.
Comment 7: Table 1 is not self-explaining. You have to interpret the table while reading.
Response: 7: Table was corrected and enhanced his quality, uniformity and self-explanation.
Comment 8: Figure 3 can substitute Table 1. In accordance with this figure, it is possible to state the most frequent boundary conditions occurring when the system disengaged.
Response: 8: This mistake has been corrected. Table 1 was misquoted, so the text has been edited > Table 1 and Fig.3 refer to other studies with different results.
Comment 9: You cannot deal with autonomous vehicles generally when assessing specific disengagements. You have to create clusters from L2 vehicles to L5 vehicles. After this clustering process, you can create manufacturer clusters according to each type of vehicle.
Response: 9: Thank you for this insightful suggestion. Unfortunately, it is not possible to disaggregate our results by level of automation because the California DMV’s Annual Report of Autonomous Vehicle Disengagement (Form O 311R) does not include any explicit classification of automation level (e.g., SAE Levels 3–5). The only autonomy‑related field reported is a binary indicator of whether the vehicle “is capable of operating without a driver,” which does not distinguish between distinct levels of automated driving functionality. Moreover, there is no standardized metadata in the disengagement reports describing Operational Design Domains, software versions, or system configurations that could serve as proxies for automation level. Consequently, the available dataset does not permit stratification of disengagement frequency or distance‑to‑disengagement metrics according to automation level. We have therefore focused our analysis on aggregate trends across all permitted vehicles and have highlighted this reporting limitation as a key area for future research and policy development.
Comment 10: Figure 6 is not useful. You could use kilometers and disengagements, not cumulating them. Currently it is not explaining and the regression line is useless. Takeovers or disengagements?
Response: 10: The graph has been replaced and the corresponding text with a description has been added.
Comment 11: Same corrections for figure 7. Is it possible that the two figures show the same distribution having different variables? This means that the variables are correlated. More statistical analysis should be done.
Response: 11: The graph has been replaced and the corresponding text with a description has been added.
Comment 12: Figures 8-10 can be useful but as they are presented, they do not contribute to the improvement of manuscript quality.
Response: 12: The relevant text has been supplemented with additional statistical and other data to provide context for the reader.
Comment 13: What does the safety driver disengagement mean? The need for driver to be more reactive than the system, the fear of the driver that the system does not perform well its task and then react before anything can be done by the system? There are several flaws in this discussion.
Response: 13: Our analysis of disengagements (i.e., deactivations of the autonomous mode) distinguishes between two primary categories—those in which the autonomous system itself elects to deactivate and those in which the safety driver takes over control. The purpose of this categorization is not to imply that the driver is compelled to be excessively reactive or is acting out of fear due to the system’s insufficient performance. Rather, it serves to quantify the shifting responsibility for safety—an indicator of how the locus of control transfers from human to machine as technology advances. As stated in the text, specific causes (e.g., perception errors, planning failures, technical faults, adverse weather, or interactions with other road users) ultimately manifest in one of two outcomes: either the system autonomously deactivates the autonomous mode, or the safety driver intervenes manually. In this way, we obtain a clear and directly comparable metric that serves as an indicator of the maturity and progress of the autonomous system. From this perspective, a “disengagement initiated by the safety driver” does not automatically imply that the driver is in a state of constant fear or excessive reactivity. Instead, it reflects the fact that, during development and testing, there remain situations where the human factor plays a crucial role in maintaining safety, particularly under complex or unforeseen conditions.
Comment 14: You mentioned different types of disengagements, why are there only two in table 2?
Response: 14: Table 2 intentionally presents only two categories of disengagement—system‑initiated versus driver‑initiated—because its purpose is not to enumerate every technical or environmental trigger, but rather to quantify the evolving locus of control between the autonomous driving system and its human operator. In practice, all specific causes of disengagement (e.g., perception errors, planning failures, hardware or software faults, adverse weather, or interactions with other road users) ultimately result in one of two outcomes: the system autonomously deciding to deactivate itself or the safety driver taking manual control. By distilling disengagements to these two initiator categories, Table 2 provides a clear, directly comparable metric—the ratio of system‑initiated to driver‑initiated interventions—that serves as a concise indicator of ADS maturity and autonomy progression over time. More granular cause‑based classifications are reported separately (see Tables 1 and 3), where the full range of disengagement triggers is analyzed; isolating initiator type in Table 2 therefore enables a focused, longitudinal assessment of how responsibility for safety transitions from human to machine as technology advances.
Comment 15: Bullet point list below the table is not adequate without any line to introduce.
Response: 15: Corrected - first bullepoint was introduction line. Incorectly it was listed as a bulletpoint. Text was also suplemented.
Comment 16: Can you divide the results according to the level of automation?
Response: 16: Thank you for this insightful suggestion. Unfortunately, it is not possible to disaggregate our results by level of automation because the California DMV’s Annual Report of Autonomous Vehicle Disengagement (Form O 311R) does not include any explicit classification of automation level (e.g., SAE Levels 3–5). The only autonomy‑related field reported is a binary indicator of whether the vehicle “is capable of operating without a driver,” which does not distinguish between distinct levels of automated driving functionality. Moreover, there is no standardized metadata in the disengagement reports describing Operational Design Domains, software versions, or system configurations that could serve as proxies for automation level. Consequently, the available dataset does not permit stratification of disengagement frequency or distance‑to‑disengagement metrics according to automation level. We have therefore focused our analysis on aggregate trends across all permitted vehicles and have highlighted this reporting limitation as a key area for future research and policy development.
Comment 17: You can use up to the second degree of polynomial regression to be consistent. Greater rates are undoubtedly forced. Statistically and mathematically, how can you represent a fourth polynomial curve? Which are the most recurrent events and phenomena that you can reproduce by a sixth-fifth grade polynomial curve?
Response: 17: Explained in detail in Fig. 11 > . In our analysis, we opted to use a fourth-degree polynomial in order to capture the potentially complex nonlinear dynamics in the evolution of the ratio of sys-tem-initiated to driver-initiated disengagements. The higher-degree polynomial ena-bled us to better describe the abrupt changes observed in the data, particularly in 2021 and 2022. Such pronounced fluctuations may occur as a result of the introduction of new software versions, changes in testing protocols, or the influence of markedly different operating conditions. Although a lower-degree polynomial yielded a simpler curve with a single peak (approximately in mid-2020), it did not accurately capture the sudden local maxima and minima evident from the limited number of data points. When fitting the data with a quadratic curve, we observed a relatively low coefficient of determination (R² = 0.358). While this curve indicates a general trend—a gradual increase until mid-2020 followed by a subsequent decrease—it fails to reflect the finer fluctuations that occurred, especially towards the end of the observed period. A similar approach was adopted in subsequent analyses, in one instance even employing a sixth-degree polynomial.
Comment 18: L. 340 - 355 Why did not investigate these other factors? Curves are similar, you have to choose the most representative ones. Including all the curves make the figure not significantly different and not remarkable. Moreover, very poor mathematical and statistical approaches are used, with very low correlations that do not explain the variability of the phenomenon.
Response: 18: Another chart (Figure 16) has been added. Thank you for your constructive criticism, which points out the shortcomings of our analysis. We agree that neglecting other factors (e.g., weather, time of day, road type, or traffic density) may lead to an inadequate explanation of variability in disconnections. In our current analysis, we were limited by the available data from disconnection reports, which do not include these variables, as we explicitly stated in our paper. Moreover, we took this approach because we had relatively few observations (only seven data points) and we believe that any additional variables would quickly reduce the reliability of the model. For future research, we suggest expanding the data set to include these variables or using alternative data sources that would allow for more comprehensive modeling.
Comment 19: Discussing results should mean that you have to assess the novelties of your results, comparing them with existing ones and improving the discussion with meaningful insights. This is a short report of the results.
Response: 19: Discussion section was substantially revised and modified.
Comment 20: Conclusion does not conclude properly.
Response: 20: Conclusion section was substantially revised and modified.
Comment 21: Citing just 15 fonts for one of the most discussed topic in the field of transportation safety is not adequate.
Response: 21: The references were supplemented by another 35 references to reach the final number of 50 references.
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI really appreciate the effort to reply to all the addressed suggestions. The current version of the manuscript provides a scientific insight into the reports utility to assess the evolution of AVs. It is increadibily important the result that the disengagement increasese by km travelled, meaning that the use of technology becomes less consistent throught time (i.e. cumulated km travelled). Figure 11, in my opinion can be discharged, even with the new explanation does not add anything new. It could be removed and one sentence about the great fluctuations over time can summarize better this phenomenon. Because the points are not correlated.
Author Response
Comment 1: Figure 11, in my opinion can be discharged, even with the new explanation does not add anything new. It could be removed and one sentence about the great fluctuations over time can summarize better this phenomenon. Because the points are not correlated.
Response 1: Dear Reviewer, thank you for your insight. We essentialy agree with your comments therefore the manuscript has been revised in accordance with the your comments and following changes were made: Figure 11 and the corresponding text (lines 493–508) have been removed, and new text requested by the reviewer has been added (lines 494–499). Additionally, the numbering of all subsequent figures has been adjusted to account for the removal of Figure 11.
Author Response File: Author Response.docx