acceptodds
Under review as a conference paper at ICLR 2027

SafeCoDriver: Post-Intervention Safety with Cooperative Perception

Abstract

Compared with single-vehicle perception, cooperative perception provides additional information about distant or occluded traffic objects, supporting earlier identification of potential conflicts. To respond to these potential conflicts, safety interventions modify a motion planner's nominal trajectory to avoid hazards. Existing methods guide these modifications through safety rules, safety constraints, and trajectory predictions to address potential hazards. However, the modifications themselves alter traffic interactions, which can introduce new conflicts or cause unnecessary slowing or stagnation. This motivates assessing post-intervention safety, i.e., whether intervention addresses hazards without introducing new collisions or stagnation. To address this need, this paper proposes SafeCoDriver, which combines the ego vehicle's onboard observations and cooperative observations to evaluate proposed interventions before execution. Trajectory Conflict Resolution (TCR) refines the complete trajectory, Multidirectional Threat Assessment (MTA) assesses surrounding threats, and Long-Horizon Feasibility Assessment (LHFA) checks predicted collisions and future lane occupancy. By integrating these components and decoupling geometric correction from the warning output, SafeCoDriver accounts for post-intervention safety. Closed-loop experiments in SUMO and open-loop evaluations on DeepAccident show that SafeCoDriver improves safety and reduces stagnation. It achieves the lowest collision rate in SUMO and the lowest waypoint exposure on DeepAccident among the evaluated safety baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Traffic Safety, Collaborative Driving

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wHkmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SafeCoDriver, a post-intervention safety assessment framework for cooperative perception in autonomous driving. It integrates three components, i.e., Trajectory Conflict Resolution, Multidirectional Threat Assessment, and Long-Horizon Feasibility Assessment, to refine planned safety interventions before execution, avoiding secondary collisions or traffic stagnation introduced by the intervention itself. The method is validated on closed-loop SUMO simulations and the open-loop DeepAccident dataset, showing improved safety and maintained mobility across multiple baselines.

Critical strengths

(1) It has a well-justified problem framing. The paper uses matched controlled experiments to clearly show that standard safety interventions can induce new collisions or stagnation in scenarios that are safe without intervention. This is an important, under-studied problem in cooperative driving. (2) It has a comprehensive empirical validation. The evaluation spans closed-loop simulation, open-loop real-world datasets, component ablation, and robustness tests under sensor noise and communication delay. The results seem to consistent and strongly support the claimed benefits.

Critical weaknesses

(1) The core technical components (beam-search trajectory correction, TTC-based threat scoring, horizon collision checking) are mostly adapted from existing methods. The work is more of a system-level integration than a fundamental advance. Thus, the novelty of this work can be further improved. (2) Surrounding vehicles are modeled as non-reactive constant-velocity agents. The paper does not evaluate how other road users’ responses to the ego vehicle’s braking or lane change would affect post-intervention safety, which weakens the real-world validity.

Questions for authors

(1) Did you test the framework with reactive surrounding driver models (e.g. IDM car-following)? Would more realistic interactive traffic change the main collision and stagnation results? (2) The design decouples trajectory correction from warning output. How would this two-branch structure map to real in-vehicle warning logic, and is there any validation of the warning performance beyond the geometric check?

4 comments
AuthorsOpenReview ↗

1/2 Response to Reviewer wHkm

We sincerely thank you for your careful reading of our paper and for your constructive comments. Your questions have guided our additional experiments and helped us clarify the method and experimental results. We address each concern below and hope that the additional evidence and explanations will help you reassess our work.

Response 1. Contribution

Reviewer comment (Weakness 1)

The core technical components (beam-search trajectory correction, TTC-based threat scoring, horizon collision checking) are mostly adapted from existing methods. The work is more of a system-level integration than a fundamental advance. Thus, the novelty of this work can be further improved.

Response

Thank you for asking us to clarify the contribution. We believe that the problem addressed in this work deserves careful investigation. A safety intervention changes ego motion and can create conflicts with surrounding traffic or prevent continued passage. Assessing these consequences is important because resolving the initial hazard alone does not establish the safety of the resulting maneuver.

SafeCoDriver explores a solution to this problem through cooperative observations and candidate intervention assessment. A suitable cooperative vehicle reveals relevant road users occluded from ego. These observations inform trajectory correction and the assessment of directional threats, future conflicts, and lane blockage before action selection. The ablation studies examine how these choices affect waypoint exposure, collisions, and mobility. The following responsive-traffic experiments demonstrate the effectiveness of this approach in the evaluated settings. We will make this problem-specific design and the role of cooperative observations clearer, with the observation assumptions and family-level results in the Appendix.

ACR is a further contribution. It measures collisions introduced into matched episodes that were collision-free under nominal driving. Reporting ACR alongside overall collisions, completion, and Progress distinguishes added risk from unresolved hazards and collision avoidance achieved through stagnation. This evaluation can be applied to other safety interventions.

Response 2. Responsive-traffic experiments

Reviewer comment (Weakness 2 / Question 1)

Surrounding vehicles are modeled as non-reactive constant-velocity agents. The paper does not evaluate how other road users’ responses to the ego vehicle’s braking or lane change would affect post-intervention safety, which weakens the real-world validity.

Did you test the framework with reactive surrounding driver models (e.g. IDM car-following)? Would more realistic interactive traffic change the main collision and stagnation results?

Response

We evaluated all nine methods across eight scenario families in responsive traffic based on the Intelligent Driver Model (IDM). The methods use the same scenario settings and acceleration and braking limits. Each method is evaluated on 1,200 matched episodes. Secondary counts episodes with a later ego collision involving an object absent from the first collision step, using all episodes as the denominator.

Table R1. Supplementary comparison under responsive traffic.

MethodCollision rate (%)Secondary collision rate (%)Completion rate (%)Progress (m)WPC (%)
No safety intervention75.00%12.33%25.00%28.882.347%
RSS adaptation74.83%0.00%24.58%19.102.261%
CBF-SC19.25%2.92%62.75%82.003.152%
DPCBF13.92%0.00%41.67%54.760.987%
UniE2EV2X-SC59.33%0.08%39.75%42.600.435%
MAP-SC22.17%1.33%72.08%123.800.047%
RiskMM-SC75.00%12.33%25.00%29.012.258%
Joint MPC42.42%27.92%57.58%106.150.033%
SafeCoDriver4.33%1.83%95.67%165.070.155%

SafeCoDriver records 52/1,200 collisions (4.33%), 95.67% completion, and 165.07 m mean Progress. It has the lowest collision rate and the highest completion rate and mean Progress in this comparison. These results show that SafeCoDriver remains effective in the evaluated responsive traffic.

AuthorsOpenReview ↗

2/2 Response to Reviewer wHkm

Response 3. Role of warning and evidence for decoupling

Reviewer comment (Question 2)

The design decouples trajectory correction from warning output. How would this two-branch structure map to real in-vehicle warning logic, and is there any validation of the warning performance beyond the geometric check?

Response

The warning is meant to alert the driver to risks. It uses a learned collision-risk score, nominal-trajectory geometric conflicts, and MTA-filtered TTC to identify risky situations and imminent threats. Cooperative observations make relevant occluded road users available to this assessment. An advisory output could help a driver recognize hazards beyond the onboard view and understand when greater attention is needed.

Even when no warning is issued, the system continues to check risks and adjust the vehicle's actions. To test this choice, we compare Full with a warning-gated variant that permits correction only when warning is active, using the same warning model and signal. The comparisons use 1,200 matched responsive episodes and 400 matched joint-input-error episodes.

Table R2. Warning-decoupling ablation.

VariantResponsive collision rate (%)Joint-error collision rate (%)
Full, decoupled4.33%3.25%
Without warning decoupling13.42%5.25%

Keeping correction independent of warning reduces collisions in both settings. This supports decoupling advisory activation from the availability of corrective action.

Dear Reviewer,

As the discussion period draws to a close, we would like to kindly follow up on our response to your review. We would appreciate hearing whether our clarifications have addressed your concerns, and whether there are any remaining points we could help clarify.

Thank you again for your time and thoughtful feedback.

Best regards, The Authors

Reviewer_wHkmOpenReview ↗

Comment

Thanks for the responses. Overall, the responsive-traffic and warning-decoupling experiments address my main questions with concrete evidence. I keep my rating at 3 (weak accept), while noting that the components are still adapted from existing work and that real-road validation remains a limitation.

Reviewer_kDvYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SafeCoDriver, a safety layer on top of a motion planner that uses onboard and cooperative perception to assess whether driving interventions introduce new collisions or stagnation. It combines trajectory correction, multidirectional threat assessment, and long-horizon evaluation of candidate driving actions, with an independent warning branch. The authors report 0% collisions and 100% completion across 1,200 closed-loop SUMO episodes, and the lowest waypoint exposure among seven baselines on DeepAccident.

Critical strengths

The paper addresses a relevant problem: safety interventions can themselves introduce new collisions or stagnation. The matched-episode analysis, which counts collisions added in episodes that were safe without intervention, is a useful way to quantify this. It is consequential because it offers a reusable diagnostic for evaluating any safety intervention, independent of the proposed method, and is the main positive factor in my assessment.

Critical weaknesses
  1. The method predicts other vehicles with constant velocity, assuming they do not react to the ego vehicle, and the SUMO benchmark appears to lack responsive traffic. This conflicts with the paper's motivation that interventions change traffic interactions, and likely favors the method, since constant-velocity prediction is nearly exact for non-reactive traffic. The results therefore say little about the realistic case the paper motivates.
  2. The evaluation appears saturated: the method and several ablations reach 0% collisions and 100% completion, ablations repeat identical values, and some baselines behave anomalously. Since the benchmark cannot distinguish the method from its own ablations, the claimed advantages are not well supported.
Questions for authors
  1. Do SUMO background vehicles react to the ego vehicle, and how closely does their motion match the constant-velocity prediction? How does SafeCoDriver perform when surrounding traffic responds to its interventions? Results under responsive traffic would most affect my assessment.
  2. The conclusion mentions collisions under combined position errors and communication delays, which Table 7 does not show. Under which conditions does SafeCoDriver collide, and how do the method and its ablations compare in those settings?
4 comments
AuthorsOpenReview ↗

1/3 Response to Reviewer kDvY

We sincerely thank you for your careful reading of our paper and for your constructive comments. Your questions have guided our additional experiments and helped us clarify the method and experimental results. We address each concern below and hope that the additional evidence and explanations will help you reassess our work.

Response 1. Responsive traffic and confusion of the original results

Reviewer comment (Weakness 1 / Weakness 2 / Question 1)

The method predicts other vehicles with constant velocity, assuming they do not react to the ego vehicle, and the SUMO benchmark appears to lack responsive traffic. This conflicts with the paper's motivation that interventions change traffic interactions, and likely favors the method, since constant-velocity prediction is nearly exact for non-reactive traffic. The results therefore say little about the realistic case the paper motivates.

The evaluation appears saturated: the method and several ablations reach 0% collisions and 100% completion, ablations repeat identical values, and some baselines behave anomalously.

Do SUMO background vehicles react to the ego vehicle, and how closely does their motion match the constant-velocity prediction? How does SafeCoDriver perform when surrounding traffic responds to its interventions? Results under responsive traffic would most affect my assessment.

Response

We evaluated all nine methods across eight scenario families in responsive traffic based on the Intelligent Driver Model (IDM). The methods use the same scenario settings and acceleration and braking limits. Each method is evaluated on 1,200 matched episodes. Secondary counts episodes with a later ego collision involving an object absent from the first collision step, using all episodes as the denominator.

Table R1. Supplementary comparison under responsive traffic.

MethodCollision rate (%)Secondary collision rate (%)Completion rate (%)Progress (m)WPC (%)
No safety intervention75.00%12.33%25.00%28.882.347%
RSS adaptation74.83%0.00%24.58%19.102.261%
CBF-SC19.25%2.92%62.75%82.003.152%
DPCBF13.92%0.00%41.67%54.760.987%
UniE2EV2X-SC59.33%0.08%39.75%42.600.435%
MAP-SC22.17%1.33%72.08%123.800.047%
RiskMM-SC75.00%12.33%25.00%29.012.258%
Joint MPC42.42%27.92%57.58%106.150.033%
SafeCoDriver4.33%1.83%95.67%165.070.155%

SafeCoDriver records 52/1,200 collisions (4.33%), 95.67% completion, and 165.07 m mean Progress. It has the lowest collision rate and the highest completion rate and mean Progress in this comparison. These results show that SafeCoDriver remains effective in the evaluated responsive traffic.

AuthorsOpenReview ↗

2/3 Response to Reviewer kDvY

Response 2. Component ablations and the contribution of TCR

Reviewer comment (Weakness 2 / Question 2)

Since the benchmark cannot distinguish the method from its own ablations, the claimed advantages are not well supported.

Under which conditions does SafeCoDriver collide, and how do the method and its ablations compare in those settings?

Response

We compare the component controls under responsive traffic and joint cooperative-input errors. The table reports collision rates and secondary collision rates. Secondary uses the definition in Table R1. We evaluated each variant on 1,200 matched episodes under responsive traffic and 400 matched episodes under joint input errors, following the protocol in Appendix D.1.

Table R2. Ablation results under responsive traffic and joint cooperative-input errors.

VariantResponsive collision rate (%)Responsive secondary collision rate (%)Joint-error collision rate (%)
Full4.33%1.83%3.25%
Without TCR4.33%1.83%3.25%
Without MTA6.17%1.25%5.00%
Without LHFA72.25%14.58%52.25%
Without warning decoupling13.42%2.00%5.25%

Table R3. TCR exposure based on recorded future trajectories on DeepAccident.

Variantexposure
Full2.24%
Without TCR3.13%

In the supplementary offline evaluation, waypoint exposure against recorded future trajectories is 2.24% for Full and 3.13% without TCR. Recorded futures are used only as the scoring reference and are not available to the method when proposals are generated. This supports TCR's role in improving waypoint clearance. The executed actions and collision rates are unchanged by removing TCR in the responsive experiment. The action-level benefits are reflected in the MTA, LHFA, and warning-decoupling comparisons. The ablations show that MTA and LHFA contribute to safer action selection: removing them increases the responsive collision rate from 4.33% to 6.17% and 72.25%, respectively. Keeping correction independent of warning also improves safety, with a collision rate of 4.33% compared with 13.42% when correction is gated by warning. The joint-input-error results show the same trend.

AuthorsOpenReview ↗

3/3 Response to Reviewer kDvY

Response 3. Table 7, Table 25, and collision conditions

Reviewer comment (Question 2)

The conclusion mentions collisions under combined position errors and communication delays, which Table 7 does not show. Under which conditions does SafeCoDriver collide?

Response

The collisions mentioned in the conclusion refer to the 13/400 episodes (3.25%) in Table 25. We will clarify the scope of both experiments and their corresponding references in Section 4.6 and the conclusion.

The SUMO results in Table 7 report a targeted robustness test under combined cooperative-input errors with a responsive following vehicle, covering five vehicle-interaction scenario families (S1, S3, S4, S7, and S8). These families were selected because they are designed around front-rear vehicle interactions, lane changes, or traffic queues, making them representative settings for evaluating interactions with a responsive following vehicle. Appendix D.1, Table 25 evaluates individual and combined cooperative-input errors across all eight scenario families. All 13 collisions occur in S5, indicating that this blind-spot crossing scenario is more sensitive to combined input errors in cooperative observations and depends more strongly on their accuracy and timeliness.

We examined all 13 failed scenarios under all four input conditions using 52 diagnostic replays. In six joint-error cases, ego issues no below-cruise speed request, remains at 12 m/s at 2.3 s, and first records a collision at 2.7 s. In the other seven cases, ego begins requesting slower speeds at 1.3 s, but the resulting speed response differs from the delay; the first collision is reported at 3.5 s. All first collision objects are the crossing pedestrian. For example, in seed 6003, the combined stale and biased shared information displaces the pedestrian laterally by approximately 1.97 m relative to its current position at 2.3 s. Ego still travels at 12 m/s, compared with 4.92 m/s with delay alone. In seed 6000, ego does slow, but its speed at the same instant is 6.64 m/s versus 4.92 m/s with delay alone. These comparisons show how joint input errors alter conflict geometry and speed response in an occluded crossing. The same 13 scenarios are collision-free with clean input or either error alone.

The separate supplementary responsive experiment in Table R1 records 52/1,200 collisions (4.33%). All occur in S3 or S4, with the following distribution by first collision object.

Table R5. Failure distribution in the supplementary responsive experiment.

ScenarioCollision episodesFirst collision object
S3 narrow-corridor interaction9Adjacent-lane vehicle
S4 cut-in with blocked escape lane23Slow vehicle cutting in ahead
S4 cut-in with blocked escape lane20Stationary obstacle in the escape lane

Representative replays illustrate the remaining failure conditions. In S3, ego enters the lane of a neighboring vehicle that continues braking; a collision follows at 1.4 s. In S4, a slow vehicle cuts in at approximately 1.8-1.9 s, followed by an ego braking and lane-change request and a collision at 2.0 s. In another S4 case, ego selects the adjacent lane containing a stationary obstacle. At 1.9 s, its speed is 11.62 m/s and the remaining longitudinal gap is 10.41 m, shorter than the ideal stopping distance of 11.25 m at the shared maximum deceleration of 6 m/s². The first collision occurs at 3.4 s. These cases motivate improved assessment of uncertain observations, changing traffic, and achievable braking.

Dear Reviewer,

As the discussion period draws to a close, we would like to kindly follow up on our response to your review. We would appreciate hearing whether our clarifications have addressed your concerns, and whether there are any remaining points we could help clarify.

Thank you again for your time and thoughtful feedback.

Best regards, The Authors

Reviewer_jNpZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses post-intervention safety, where safety interventions can introduce new collisions or cause stagnation. Using onboard and cooperative observations, the method combines four components: TCR for waypoint correction, MTA for directional threat assessment, LHFA for 5-second action checking with lexicographic selection, and a learned warning branch that is decoupled from correction. The authors test the method in closed-loop SUMO over 1,200 episodes from eight scenario families and report 0% collisions and 100% completion. In an open-loop test on 104 DeepAccident scenarios, WPC drops from 3.767% to 0.380%. The paper also proposes ACR, a metric that counts collisions added by intervention.

Critical strengths

S1. Problem formulation and ACR

ACR counts intervention-induced collisions using matched episodes that differ only in whether an intervention is applied. The paper reports completion and stagnation along with collisions. This shows when a method avoids collisions only by stopping. Out of 1,200 episodes, 902 for RSS and 840 for DPCBF are collision-free but fail to complete.

Critical weaknesses

W1. Narrow evidence for the central claim

The central claim is that interventions create risk by changing traffic interactions. However, the experiments use constant-velocity, non-reactive prediction (Eq. 1), a straight-line nominal trajectory (App. E.1), and prescribed rear-threat vehicles (App. B.6). The responsive-follower experiment covers only five families and two baselines, DPCBF and Joint MPC. The claim that every baseline introduces new collisions is based on 154 episodes: 150 in S3 and four in S1. For RSS, the evidence comes only from the four episodes in S1. Changes in interaction are the paper's main claim. Unless the authors show that this claim holds more generally, the contribution is much smaller.

W2. Lack of independent safety validation on DeepAccident

WPC and PFE are computed against the same constant-velocity predictions that TCR optimizes against. Because the TCR margin of at least 2.5 m exceeds the WPC threshold of 2 m, WPC is effectively the residual left by TCR. The colliding-vehicle IDs and ground-truth future trajectories provided by DeepAccident are not used as the reference. MTA and LHFA are not reflected in this metric. The inputs are ground-truth annotations, and both benchmarks are simulated.

Questions for authors

Q1. Clarification on the SUMO configuration

Please report the SUMO version, the collision-related settings, the ego vehicle's speedMode and laneChangeMode, and how the threat vehicles are controlled. Two results are hard to interpret without this information. In S5, “No safety intervention” records 150/150 collisions but 0/150 body contacts (Table 8). Are collisions registered when the gap falls below minGap rather than when physical contact occurs? In S8, the nominal straight trajectory (App. E.1) ends in terminal stagnation in all 150 episodes without any collision (Table 9). Why does the vehicle stop? For example, does the ego vehicle retain SUMO's default car-following safety checks?

Q2. Clarification on onboard and cooperative observations

How do you construct ego visibility (including occlusion and sensing range) and cooperative observations in SUMO? For example, which agents share observations, from where, and how is occlusion computed? Could you report per-family results for the onboard-only variant in Table 3? Its 12.50% collision rate corresponds to exactly 150 of the 1,200 episodes. These answers would help clarify what the cooperative-perception gain in Table 3 comes from and whether it holds across families or comes from a single family.

Q3. Additional baseline evidence under the responsive follower condition

The conclusion lists evaluation with responsive traffic as a limitation, and Sec. 4.6 explains why only DPCBF and Joint MPC were included in the responsive-follower experiment. However, the RSS entries in Table 1, which Sec. 4.2 uses to explain the interaction in Fig. 1a, come from rear-end collisions with a prescribed follower. A small additional check would therefore help. Could you evaluate at least RSS under the responsive-follower setting in App. B.6? In the four RSS rear-end collisions in S1 (App. C), what deceleration did the ego vehicle apply? Was it within the maximum braking assumed by RSS's safe longitudinal distance? If these collisions disappear when the follower responds, the claim that every evaluated baseline introduces new collisions would no longer hold for RSS. This would also weaken the evidence behind Fig. 1a.

8 comments
AuthorsOpenReview ↗

1/5 Response to Reviewer jNpZ

We sincerely thank you for the time and care you devoted to reviewing our paper and supplementary material. We particularly appreciate your close examination of the experimental settings and the results across the main paper and appendices. Your detailed questions and constructive comments have guided our additional experiments and helped us clarify the implementation and the scope of our conclusions. We address each concern below and will include the supporting experimental and implementation details in the supplement.

Response 1. Responsive traffic and prediction mismatch

Reviewer comment (Weakness 1)

The central claim is that interventions create risk by changing traffic interactions. However, the experiments use constant-velocity, non-reactive prediction (Eq. 1), a straight-line nominal trajectory (App. E.1), and prescribed rear-threat vehicles (App. B.6). The responsive-follower experiment covers only five families and two baselines, DPCBF and Joint MPC.

Response

We extended the responsive-traffic experiment to all nine methods and all eight scenario families, using IDM (Intelligent Driver Model) traffic. All methods use the same scenario and execution settings. SafeCoDriver retains its constant-velocity predictor and records a 4.33% collision rate and 95.67% completion, which are the lowest collision rate and highest completion rate among all evaluated methods, respectively. The full comparison is in Table R1.

Table R1. Supplementary comparison under responsive traffic.

MethodCollision rate (%)Secondary collision rate (%)Completion rate (%)Progress (m)WPC (%)
No safety intervention75.00%12.33%25.00%28.882.347%
RSS adaptation74.83%0.00%24.58%19.102.261%
CBF-SC19.25%2.92%62.75%82.003.152%
DPCBF13.92%0.00%41.67%54.760.987%
UniE2EV2X-SC59.33%0.08%39.75%42.600.435%
MAP-SC22.17%1.33%72.08%123.800.047%
RiskMM-SC75.00%12.33%25.00%29.012.258%
Joint MPC42.42%27.92%57.58%106.150.033%
SafeCoDriver4.33%1.83%95.67%165.070.155%

Among the nine methods in Table R1, SafeCoDriver has the lowest collision rate and the highest completion rate and mean Progress.

Each method is evaluated on 1,200 matched episodes. Secondary counts episodes with a later ego collision involving an object absent from the first collision step, using all episodes as the denominator. The full protocol is provided in the Appendix.

AuthorsOpenReview ↗

2/5 Response to Reviewer jNpZ

Response 2. RSS and the evidence for added collisions

Reviewer comment (Weakness 1 / Question 3)

The conclusion lists evaluation with responsive traffic as a limitation, and Sec. 4.6 explains why only DPCBF and Joint MPC were included in the responsive-follower experiment. However, the RSS entries in Table 1, which Sec. 4.2 uses to explain the interaction in Fig. 1a, come from rear-end collisions with a prescribed follower. A small additional check would therefore help. Could you evaluate at least RSS under the responsive-follower setting in App. B.6? In the four RSS rear-end collisions in S1 (App. C), what deceleration did the ego vehicle apply? Was it within the maximum braking assumed by RSS's safe longitudinal distance? If these collisions disappear when the follower responds, the claim that every evaluated baseline introduces new collisions would no longer hold for RSS. This would also weaken the evidence behind Fig. 1a.

The claim that every baseline introduces new collisions is based on 154 episodes: 150 in S3 and four in S1. For RSS, the evidence comes only from the four episodes in S1. Changes in interaction are the paper's main claim. Unless the authors show that this claim holds more generally, the contribution is much smaller.

Response

We additionally evaluated RSS using the same scenes, seeds, and execution settings as Appendix B.6, with 200 matched episodes per condition, comparing runs with and without a responsive IDM follower. All 19 RSS collision episodes involve the responsive follower, while the corresponding runs without safety intervention are collision-free. Removing the follower eliminates these RSS collisions. These results show that collisions can still occur when the follower responds and demonstrate the need to assess an intervention's effects on following traffic.

Table R2. RSS evaluation under the responsive-follower setting in App. B.6.

MethodWith follower: collision episodes / totalWithout follower: collision episodes / totalWith follower, matched subset of RSS collision episodes / total RSS collision episodes
No safety intervention40/20040/2000/19
RSS adaptation19/2000/20019/19
SafeCoDriver0/2000/2000/19

To ensure a fair comparison, all methods should use the same vehicle capabilities and acceleration and braking limits. These shared physical constraints provide the basis for comparing the implemented methods; RSS's safety guarantees also depend on its stated assumptions. We will detail the vehicle capabilities and execution limits in the Appendix.

AuthorsOpenReview ↗

3/5 Response to Reviewer jNpZ

We would like to clarify that the added-collision analysis in Table 1 includes all 310 episodes that are collision-free without safety intervention. The overall ACR also uses these 310 episodes as its denominator. Of these, 154 are collision-free and not stagnant: four in S1 and 150 in S3. The other 156 are collision-free but stagnant: six in S7 and 150 in S8. The full analysis therefore covers S1, S3, S7, and S8. All seven comparison safety methods introduce collisions in the nominally completed group. CBF-SC, MAP-SC, RiskMM-SC, and Joint MPC also introduce collisions in the nominally stagnant group. The counts and scenario families are listed below.

Table R3. Added collisions in completed and stagnant episodes.

MethodAdded-collision episodes / eligible episodes (completed group)Added-collision episodes / eligible episodes (stagnant group)
RSS4/154 (S1: 4)0/156
CBF-SC11/154 (S3: 11)123/156 (S8: 123)
DPCBF1/154 (S3: 1)0/156
UniE2EV2X-SC10/154 (S1: 4; S3: 6)0/156
MAP-SC2/154 (S3: 2)41/156 (S8: 41)
RiskMM-SC139/154 (S3: 139)156/156 (S7: 6; S8: 150)
Joint MPC2/154 (S3: 2)28/156 (S8: 28)
SafeCoDriver0/1540/156

These matched results show that, in the evaluated scenarios, a safety intervention can cause a collision in an episode that was collision-free without intervention.

We acknowledge that the RSS added-collision evidence covers only S1. In this simulation sample, RSS adds collisions in four of the 310 episodes, which is a small observed proportion. This does not establish that intervention-added collisions are absent from real traffic. Even infrequent failures matter for driving safety, and their causes and prevention remain worth studying.

Prior research supports examining how a maneuver affects following traffic. In a four-vehicle braking model, Touran et al. (1999) found that autonomous intelligent cruise control could reduce collision risk with the vehicle ahead while adversely affecting following vehicles under the model assumptions. Althoff and Dolan (2014) assess future occupancy conflicts across traffic participants. In police-reported crash data, Cicchino (2017) reported that the studied forward-collision-warning systems with autonomous emergency braking reduced rear-end striking involvement by 50% while rear-end struck involvement increased by 20%. Feng et al. (2023) further motivate safety validation under rare critical events. Together, these studies support checking the broader consequences of a proposed maneuver, such as a collision introduced by an intervention.

References

A. Touran, M. A. Brackstone, and M. McDonald. A collision model for safety evaluation of autonomous intelligent cruise control. Accident Analysis & Prevention, 31(5):567-578, 1999. doi: 10.1016/S0001-4575(99)00013-5. URL https://doi.org/10.1016/S0001-4575(99)00013-5.

Matthias Althoff and John M. Dolan. Online verification of automated road vehicles using reachability analysis. IEEE Transactions on Robotics, 30(4):903-918, 2014. URL https://publications.ri.cmu.edu/online-verification-of-automated-road-vehicles-using-reachability-analysis.

Jessica B. Cicchino. Effectiveness of forward collision warning and autonomous emergency braking systems in reducing front-to-rear crash rates. Accident Analysis & Prevention, 99:142-152, 2017. doi: 10.1016/j.aap.2016.11.009. URL https://doi.org/10.1016/j.aap.2016.11.009.

Shuo Feng, Haowei Sun, Xintao Yan, Haojie Zhu, Zhengxia Zou, Shengyin Shen, and Henry X. Liu. Dense reinforcement learning for safety validation of autonomous vehicles. Nature, 615:620-627, 2023. doi: 10.1038/s41586-023-05732-2. URL https://doi.org/10.1038/s41586-023-05732-2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses to Reviewers

We thank all reviewers for their constructive feedback. Our main additions and clarifications are:

  • Responsive-traffic evaluation: All nine methods were evaluated across eight scenario families, with 1,200 episodes each. SafeCoDriver achieves the lowest collision rate (4.33%) and highest completion rate (95.67%).
  • Stronger ablations: Responsive-traffic and joint-input-error experiments clarify the contributions of MTA, LHFA, and warning decoupling, while independent exposure evaluation supports TCR’s role.
  • Greater transparency: We expand the analyses of intervention-added collisions, scenario-dependent cooperative-perception benefits, implementation details, and remaining failures.

We incorporated these additions into the revised manuscript and supplement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.