Research on Takeover Safety of Intelligent Vehicles Based on Accident Scenarios in Real-Vehicle Testing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

General Thoughts

 The study looks at an important and timely issue of takeover safety in AVs. The study's main strength is its method, which uses real-world accident scenarios to ground the experimental work. The paper is mostly well-organised and gives a clear picture of the data that was collected. Quantifying the shortest safe takeover times is a useful addition to the field.

The manuscript has a good base, but there are a few important parts that need a lot of work before it can be published. The main problems are that the literature review is too dnarrow, the results are not put into context, and the methodology is not properely organized . 

The manuscript requires significant modifications. The literature review and positioning are not enough for a scientific research paper. It talks about big ideas like driver fatigue and over-reliance, but it doesn't give a full picture of all the previous experimental (or enve simulatinmn) studies on the specific failure scenarios being looked at.

The authors should do the following to make the manuscript stronger:

Do a focused review of past studies that have looked at how well L2/L3 systems work on large-curvature curves and for e.g how well they can see static, non-standard, and irregular obstacles.

Better explain the current state of the art and clearly state what new contributions this work brings to research and industry. The paper's strength is that it uses real accident data to guide testing protocols. This should be made clear as a gap in the current literature if it is true.

 The "Discussion" section would be much better if it compared the paper's results to those of other studies. The current discussion mostly looks at the results on their own.

When looking at the quantitative results, like the 4.12-second required takeover lead time for longitudinal control and the less-than-one-second critical takeover window for lateral control, they should be compared directly to specific values from other studies, whether they were done in real life or in a simulation, that looked at similar situations.

A major finding is that the test vehicles couldn't see overturned cars or traffic cones. You should talk about this finding in relation to other studies that have been published that show the known problems with automotive perception systems. This comparison is necessary to confirm the results and show how important they are.

Some previous research you might consider (just examples)

https://doi.org/10.1016/j.treng.2024.100264

https://doi.org/10.1016/j.vehcom.2023.100623Get rights and content

https://doi.org/10.1016/j.treng.2020.100029
https://doi.org/10.1177/0361198122114143

The "Materials and Methods" section is ok, but it would be much clearer if it had a flowchart and an introductiory sub chapter.  The research process is hard and is only shown in text, including the tasks of looking at accidents and figuring out what the data means. Please do not begin presenting scenarios without first providing an overview of the general methodology.

A few small changes and some specific questions

In the curve scenario accident, the NOA system turned off without any warning or prompt to take over. Could the authors explain if this silent disengagement is a feature or a bug in the system? Was the vehicle working outside of its Operational Design Domain ?

The study treats system warnings as the same as a Take over time TOR because there was no clear TOR given during testing. Can you  explain  if the systems that were tested are set up to send explicit TORs in any situation, or if a warning is the only way to get an alert.

The introduction correctly points out how passive fatigue affects how ready a driver is. Were the test drivers alert in a controlled way? 

chapter 3.1- The results show that Vehicle A left its lane while Vehicle B stayed in the R250M curve at 90 km/h, even though it moved up and down a lot. Can the authors say if differences in how the vehicles move (e.g  the suspension, steering systems) could explain this performance gap?

The study finds that a minimum safe takeover time of 4.12 seconds is needed for longitudinal control. Could the authors give a clearer explanation or calculation of how this specific value was found using the test results in Table 8 and the literature that was cited?

In highlights and abstract you mention the term "Human-Machine Cooperative Driving (HMCD)" It would be helpful to clearly explain this term in the introduction and connect it to the standard SAE Levels of Automation.

274 - The authors mention a "professional driver's limit takeover time" of 1.40 seconds. Please give a source or a more in-depth explanation of how this baseline was set up.
Conclusions are ok for me

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study makes valuable contributions to understanding takeover safety in intelligent vehicles through rigorous real-world testing. The investigation of large-curvature curves and static obstacles addresses critical gaps in autonomous driving research, with practical implications for system design.  

1. Testing only right curves may overlook asymmetric vehicle dynamics in left curves, Limited obstacle types (e.g., missing moving/partially occluded objects) reduces scenario coverage.  
2. Unclear number of test repetitions and driver sample size affects result reliability.  
3. Missing driver demographics (age/experience) prevents analysis of individual differences.  
4. Proposed 4.12s/1.87s thresholds need comparison with industry standards (e.g., ISO/SAE) in discussion.
5. It is suggested that the paper should address how these values account for varying driver states (e.g., distracted/fatigued) for the consideration of paper subject.

  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The theoretical innovation of the paper is somewhat lacking, as it mainly focuses on experimental research. It is advisable to see if the theoretical innovation of the paper can be improved.
2. The paper's description of specific testing methods is not detailed enough, making it difficult to replicate the experimental methods. Please clearly describe elements such as the experimental process and testing methods.
3. "It is suggested to strengthen the early warning system in autonomous driving to ensure sufficient takeover time." These suggestions should be routine ones, and the paper lacks innovation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper investigates takeover safety through real vehicle testing in two typical accident scenarios: large-curvature curves and static obstacles. This paper is interesting. The comments are as follows.

1. The introduction section is too short, more detailed section such as motivations should be further highlighted.
2. What is the key innovation of this paper?
3. What are the connections between 2.1.1 and 2.1.2?
4. There are T1 and T3 in Equation 3, where is T2?
5. More detailed comparison results should be added in this paper.
6. The paper lacks detail for potential future directions, the authors are suggested to add detailed future directions. For example, if the disturbance is considered, model predictive control and dual control can be used to tackle this type of disturbance, it is worth exploring whether the model predictive control and dual control can be used to the proposed framework if the disturbance is considered, such as “A two-layer optimal scheduling method for microgrids based on adaptive stochastic model predictive control” and “Dual control for autonomous airborne source search with Nesterov accelerated gradient descent: Algorithm and performance analysis”.
7. When conducting the actual experiment, how to obtain the data?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Comments and Suggestions for Authors

Overall comments:

Overall, this is an interesting investigation into L1 / ADAS features and their ability to handle complex scenarios, which are grounded in two real-world crash examples. However, a lack of methodological detail, and the lack of real L2 features in both vehicles make it difficult to interpret the results in terms of take-over request performance. Both vehicles appear to have one system that is truly an L1 system—ACC in Vehicle A, and lane-centering in Vehicle B, but neither have both, and aren’t Level 2, and as such the driver isn’t just monitoring the system, but is an active driver in both vehicles—at least as far as I can tell. The vehicles with more full-blown automation systems that are appropriate for each scenario—lane-keeping in Vehicle B, longitudinal control in Vehicle A—appear to handle the scenarios better when the demands meet their capabilities, as one would expect. However, it’s very confusing, because the results aren’t clearly tied to these vehicles’ different capabilities, and I’m not sure I’m interpreting it correctly. Also, without more knowledge about what systems are being compared (what makes, models, and years), it’s hard to conclude much. Significantly more methodological detail, especially vehicle capabilities—and explicitly tying those capabilities to the results—is needed.

Minor edits and typos:

Line 76: Define NOA (navigate on autopilot) the first time the initialization is used.

Specific comments:

Table 1: At what point was the automated lane change initiated / terminated—was this completed before the -5 second window (I’m assuming this is from the event data recorder)? Were there any auditory or visual warnings or messages that the vehicle was changing lanes, or that it was exiting the highway? In this scenario, does the vehicle normally automatically slow down, or does it usually simply deactivate lateral and longitudinal control?

Table 3: Did either of these vehicles have driver monitoring systems (DMS)? Were drivers required to keep their hands on the wheel during lane keeping or lane centering activation? Both of these cars sound like they have very simple L2 systems, with Vehicle A having ACC but no true lane-centering system, and Vehicle B having lane-centering but apparently not ACC, so in both cases drivers could not really take their hands off the wheel and feet off the pedals and let the car drive, as how drivers sometimes mis-use other L2 systems that do offer true lane-centering and ACC. As such, I would not expect either system to issue take-over warnings because it is very unlikely the driver is disengaged.

Table 4: In the United States, interstate highways with speed limits accommodating 120 km/h cannot have curve radii as low as 250m—this would be a fairly high g-force event for a driver, and, in rain, nearing the limits of tire traction. Why was this speed / radius combination chosen—what kind of scenario does it represent?

Equation 1: I’m confused as to how this computes the distance between the front left wheel and the left lane line, as l and w appear to be constants (the lane width and host vehicle width, respectively). Is this the threshold—i.e., if the center of the host vehicle moves left by this much then it’s crossed the left lane?

Line 144-146: Since you are collecting takeover times, I assume there is a human-in-the-loop in this testing procedure—was it a member of the research team, or a professional driver, or a naïve participant? Did the same driver experience every scenario in both vehicles? Were the scenarios repeated, and were their orders randomized? Some more detail about testing would be helpful.

Line 156: My understanding of Vehicle A is that, since it doesn’t offer true lane centering (only lane departure assistance, which turns the vehicle away from a lane line as it approaches it, rather than keeping the vehicle at lane center), it was simply bouncing the vehicle between lane lines to avoid lane departure—is this correct?

Line 161-164: Did both systems accurately detect lane lines throughout the curve? Lane line detection varies considerably from vehicle to vehicle and scenario to scenario.

Line 202: Here, Vehicle B seems to only have ADAS features for alerting and emergency braking to forward collision events, but not true automation features (such as ACC), so this isn’t very surprising—forward collision warnings to stationary objects are unfortunately quite variable, and aren’t meant to be “takeover alerts” because the driver is only using conventional cruise control with some safety features, as opposed to true longitudinal control (as in Vehicle A).

Line 206: This is probably true, because Vehicle A appears to have ACC enabled which can recognize lead vehicles and adjust speed based on time headway, but may struggle with other fixed objects because in that case it is likely relying on its FCW / AEB features.

Table 8: These are really quite bad recognition performances for both vehicles, especially Vehicle B—even aftermarket FCW sensors perform better than this. Are these older vehicles?

https://www.sae.org/publications/technical-papers/content/2025-01-8677/

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you to the authors for carefully addressing all my comments. I am happy with this version and recommend it for publication

Author Response

Thank you for your reply and comments.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper can be accepted in present form.

Author Response

Thank you for your reply and comments. In this revision, we have thoroughly revised our English text and utilized MDPI's editing services for refinement. Please find the corresponding revisions/corrections highlighted in the re-submitted files.

Reviewer 4 Report

Comments and Suggestions for Authors

The authors didnot give further directions using model predict control and dual control from the two papers in comment 6, the references section is a little messy, and the readibility of this paper is poor. So the authors should take all the questions into account carefully in the revision.

Author Response

In this revision, we have thoroughly revised our English text and utilized MDPI's editing services for refinement. Please find the corresponding revisions/corrections highlighted in the re-submitted files.

Comments 1: [The authors didnot give further directions using model predict control and dual control from the two papers in comment 6, the references section is a little messy, and the readibility of this paper is poor. So the authors should take all the questions into account carefully in the revision.]

Response 1: [Thank you for pointing this out. In response to your comments, we have included the impact of predictive models and optimization methods on autonomous driving systems in our future research plans. Efficient hierarchical predictive models will enhance the overall performance of autonomous driving systems across diverse scenarios. Additionally, we have thoroughly revised the reference section to ensure greater consistency throughout the manuscript.] -line532-545

[Research in other fields, extending beyond the automotive domain, has recently proposed optimisation and control approaches that may serve as a source of inspiration for future advancements in takeover strategies. For instance, Tan et al. proposed a dual control framework with Nesterov accelerated gradient descent (DCEE-NAGD) for autonomous airborne source search, which effectively balances exploration and exploitation, thereby enhancing search efficiency and robustness[42]. In a similar vein, Hu et al. developed a two-layer optimal scheduling method for microgrids based on adaptive stochastic model predictive control. In this method, adaptive period partitioning and scenario-based constraints significantly reduced the negative impact of uncertainty on system performance[43]. The extant literature indicates that the integration of advanced predictive control and optimisation algorithms into automated driving systems has considerable potential. By leveraging such approaches, future takeover strategies can be rendered more adaptive and resilient, thereby enhancing both driver safety and system reliability in complex driving environments.]