acceptodds
Under review as a conference paper at ICLR 2027

SRTDrive: Safety-Aware Trajectory Refinement via Reference-Relative Supervision

Abstract

End-to-end planners output a single final plan that can still collide with surrounding agents, even when a small adjustment could avoid the collision. Retraining the planner changes its weights and acquired behavior, and learned candidate scorers are not trained to judge a replacement relative to the original plan. We propose SRTDrive, a post-hoc refinement framework that leaves the planner frozen, keeps its output unchanged as a reference trajectory, and learns whether to retain or replace it. A generator produces alternatives as deformations of the reference along and across its path, so that every alternative is directly comparable to the original plan. A selector then scores each candidate relative to the reference and is trained so that a replacement that increases collision at any horizon is never the target, however close it is to the logged trajectory; trajectory error only ranks equally safe candidates. At inference, the top-ranked candidate, possibly the reference itself, is output without a collision threshold. On Bench2Drive, SRTDrive raises the Driving Score of VAD-Base from 43.07 to 47.77 and reduces the routes with collisions from 95 to 86, whereas the same candidates selected by a TTC rule do not yield these gains. On nuScenes, applied to seven frozen planners without retraining, it lowers L2 error on five and the collision rate on three, with the largest gains on VAD. The code will be released at https://srtdriveiclr2027.github.io/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Safety-Aware Trajectory Refinement, Reference-Relative Supervision, Post-hoc Planner Refinement, Lexicographic Supervision, Rectified Flow

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SeU6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SRTDrive, a post-hoc trajectory refinement framework for end-to-end autonomous driving. Given the trajectory and scene features from a frozen planner, SRTDrive keeps the original trajectory as a reference candidate and generates several alternatives by applying progress, normal, and tangential deformations to it. A set-contextual selector then compares all candidates with the reference. It first excludes candidates that increase collision relative to the reference, and uses displacement error to rank equally safe candidates. The method is evaluated with seven frozen planners on nuScenes and with VAD on Bench2Drive.

Critical strengths
  1. SRTDrive provides a practical post-hoc interface that does not modify the weights of the base planner. Keeping the original plan in the candidate set also gives the selector the option to retain it when the generated alternatives are unsuitable.

  2. The Bench2Drive experiment shows a meaningful improvement with VAD as the base planner. SRTDrive improves both Driving Score and Route Completion while reducing collisions.

Critical weaknesses
  1. My major concern is the limited novelty. Trajectory refinement around an initial plan is a well-established idea, and using a rectified-flow model to generate refined trajectories appears to be a relatively direct application of an existing generative technique. The paper should position this contribution more concretely against prior refinement methods, preferably with empirical results on Bench2Drive or nuScenes.

  2. The nuScenes results do not consistently support the claimed safety benefit. Collision rate improves for only three of the seven planners. Thus, four out of seven planners receive no collision improvement, while the gains outside VAD are modest. Moreover, open-loop trajectory evaluation on nuScenes is known to be affected by ego-status shortcuts. A more informative planning benchmark such as NAVSIM would provide stronger evidence.

  3. The ablation study is insufficient for a framework containing many components. For example, the paper does not study the effects of the components in candidate generation (e.g., progress/normal/tangential plan warping, path-local and global feature fusion, and rectified-flow decoder).

  4. There are also some issues with presentation. Figure 1 provides little technical information beyond stating that the method generates and selects a refined trajectory.

Questions for authors

I don't have other questions. Please refer to the Critical Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer SeU6 (Part 1/2)

Thank you for your careful review and constructive suggestions. We agree that the differences from existing trajectory refinement methods and the contributions of individual components require clearer validation. The central contribution of SRTDrive is not Rectified Flow itself, but the use of the original trajectory as both an anchor for generating alternatives and a reference for learning retain-or-replace decisions based on relative collision outcomes.

1. Differences from existing refinement methods and the effect of the selection design

ResWorld's Future-Guided Trajectory Refinement (FGTR) [1] directly refines the initial trajectory using future BEV features. In contrast, SRTDrive preserves the original trajectory unchanged in the candidate set and uses inter-candidate relationships and reference-relative features to determine whether to retain or replace it.

The training target prioritizes avoiding increased collisions relative to the original and minimizing collisions, with L2 used to distinguish candidates within these priorities. This criterion is applied during training. At inference time, selection relies on learned scores without using ground-truth collision information or an additional collision threshold.

CLOVER [3] uses a Scorer that predicts evaluation metrics to improve the Generator and, when needed, reranks candidates according to their deviation from a reference trajectory. In contrast, SRTDrive preserves the frozen planner's original trajectory as both a candidate and the reference for training retain-or-replace decisions that prioritize avoiding collision deterioration.

To provide a direct comparison, we kept the official ResWorld model frozen and evaluated FGTR and an independently trained SRTDrive using the same initial trajectories. SRTDrive was trained separately using ResWorld features. The evaluation included 5,119 nuScenes inputs with valid future annotations.

Refinement MethodMean L2 (m) ↓Mean Collision Metric (%) ↓
Original Trajectory0.31740.0364
ResWorld FGTR0.30200.1178
SRTDrive0.31880.0364

SRTDrive achieved a lower collision metric than FGTR but a higher L2 error. Compared with the original trajectory, SRTDrive maintained the same collision metric while slightly increasing L2. Therefore, we do not interpret these results as an additional improvement over the original trajectory or as evidence of superiority over existing refinement methods in general.

We further evaluated the effect of the selection objective in a separate nuScenes experiment. Using identical initial weights, candidate sets, and training budgets (3 epochs), collision-prioritized supervision reduced replacements that introduced collisions absent from the original trajectory from 6 to 2 compared with L2-only training, while maintaining 10 cases in which original collisions were resolved.

Mean L2 increased from 0.6976 to 0.7076 m, while the mean collision metric decreased from 0.1519% to 0.1389%. This indicates that the collision reduction cannot be attributed simply to selecting trajectories closer to the ground truth. However, the reduction corresponds to only four inputs and does not establish a general safety guarantee.

We also compared the Full Selector with an additional Learned Pointwise Scorer in Bench2Drive over 220 routes. The Full Selector reported in Table 2 achieved a higher Driving Score (47.77 vs. 40.14) and fewer routes with collisions (86 vs. 108). The original VAD-Base results were 43.07 and 95, respectively.

Thus, simply adding a learned Scorer did not reproduce the improvements obtained with the full selection framework. This comparison evaluates the overall selection design and should not be interpreted as isolating the effect of any single component.

Continued in Part 2/2: NAVSIM evaluation, component analysis, and architecture clarification.

AuthorsOpenReview ↗

Response to Reviewer SeU6 (Part 2/2)

Continued from Part 1/2.

2. Applicability on NAVSIM and component analysis

On nuScenes, collision metrics improved for only 3 of the 7 evaluated planners, with relatively small improvements outside VAD.

To complement the limitations of open-loop evaluation, including concerns about the ego-status shortcut, we applied SRTDrive trained for a 4-second prediction horizon to VADv2 [2] and evaluated it on 12,146 NAVSIM v2 inputs.

The following table summarizes the Single-Stage evaluation and additional Generator ablation and replacement experiments. For these experiments, we matched the initial weights, data, number of candidates, and training budget. The Selector was also trained for the same number of epochs for each Generator configuration.

The reported values are aggregate scores of the final selected trajectories, scaled to 100 points.

ConfigurationNon-reactive ↑Reactive ↑
Original VADv282.16781.936
Full Generator82.16681.926
Without Progress82.13681.906
Without Normal82.13281.912
Without Tangential82.18081.947
Without Route-Local Features82.14981.891
Without Global Scene Features82.15081.899
Rectified Flow Replaced with Direct-Prediction Decoder82.09781.907

Although NC, DAC, and TTC increased by approximately 0.02–0.03 points, the Full Generator did not improve the aggregate score over the original VADv2.

Except for removing Tangential, the ablated or replaced configurations produced slightly lower aggregate scores than the Full Generator. However, the differences are small, and we do not claim that these results establish the necessity of each component.

In particular, no additional benefit from Tangential was observed, while the direct-prediction Decoder achieved higher Comfort.

These are Single-Stage evaluation results and do not include Two-Stage evaluation. They should also be distinguished from the Bench2Drive closed-loop setting, which involves repeated replanning.

3. Clarifying the core architecture and the distinction between training and inference

We will revise Figure 1 to explicitly distinguish three components: reference-anchored trajectory deformation, the candidate set containing the unchanged original trajectory, and the Selector using inter-candidate attention and reference-relative features.

Collision information and ground-truth annotations will be connected only to the training-target computation path. This will clarify that SRTDrive does not perform explicit collision checking or collision-based rejection at inference time.

We will also align the explanations in Figure 2, Eqs. (6)–(8), and Appendix A.6 to ensure a consistent description of the training and inference procedures.

Thank you again for helping us clarify the distinction from existing approaches and the scope of our experimental validation.

References

[1] Zhang, J., Fu, Z., Xu, Z., Dai, W., Liu, Q., and Wang, Y. ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving. ICLR, 2026.

[2] Jiang, B., Chen, S., Gao, H., Liao, B., Zhang, Q., Liu, W., and Wang, X. VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning. ICLR, 2026.

[3] Ang, S., Yang, Y., Chen, C., and Wang, Y. CLOVER: Closed-Loop Value Estimation & Ranking for End-to-End Autonomous Driving Planning. arXiv:2605.15120, 2026.

AuthorsOpenReview ↗

Additional Clarification on Novelty (Reviewer SeU6)

After posting our earlier response, we continued to reflect on your concern about novelty and felt that we should clarify the problem our work aims to address and how it differs from existing methods. We would therefore like to offer a further explanation of the motivation behind our approach.

As you noted, refining an initial trajectory or applying Rectified Flow is not, in itself, a new approach. SRTDrive aims to preserve the option to retain the original plan and learn whether to replace it with a generated alternative, without retraining or replacing the base planner. From this perspective, the distinction we wish to emphasize is using the original trajectory as a common reference for both alternative generation and learning whether to retain or replace the plan. We keep the original trajectory unchanged in the candidate set, and the selector learns this decision using inter-candidate relationships and reference-relative features. In particular, the training targets are constructed so that an alternative that worsens the collision metric relative to the original is not preferred merely because it has a lower L2 error. The comparisons with existing refinement methods and alternative selection designs presented in our earlier response were likewise intended to examine the effects of this design.

We recognize that this design does not guarantee improvements for every planner, and our results show limited gains or performance degradation for some planners. With these limitations in mind, we have clarified how our approach differs from prior work, the problem it seeks to address, and the scope of the contributions supported by our experiments in the revised manuscript. Reflecting further on your concern about novelty has helped us articulate our contribution more clearly. Thank you again for your valuable feedback.

AuthorsOpenReview ↗

Further Clarification of Our Core Contribution and Additional Experimental Evidence

Thank you very much for your careful review and thoughtful comments. We sincerely appreciate the attention you have given to the novelty and experimental validation of our work.

Your concern regarding the distinction between SRTDrive and existing trajectory refinement methods was particularly valuable to us. It encouraged us to reconsider how we communicate the central motivation of our approach and what experimental evidence is needed to support its contribution.

Following your suggestions, we have revised the related discussion and conducted several additional experiments. We would be grateful for the opportunity to briefly share the most relevant findings.

1. Clarifying the core contribution of SRTDrive

As you rightly pointed out, trajectory refinement and Rectified Flow are established techniques. We have therefore clarified that the main contribution of SRTDrive lies not in using Rectified Flow itself, but in integrating reference-anchored alternative generation with reference-relative retain-or-replace supervision.

SRTDrive preserves the original planner's trajectory unchanged as a candidate and uses it as the reference for learning whether a generated alternative should replace it. The Selector is trained to prioritize avoiding collision deterioration relative to the original trajectory, rather than simply scoring each candidate independently.

We have made this distinction more explicit in Section 2.2 and the revised architecture figures. We also added a direct comparison with ResWorld's FGTR and further clarified the differences from CLOVER (Appendix D.5).

2. Additional evidence for the proposed selection framework

To further examine whether this design provides benefits beyond conventional learned scoring, we compared the Full Selector with a Learned Pointwise Scorer using the same Generator.

In the Bench2Drive closed-loop evaluation, the Full Selector achieved a higher Driving Score (47.77 vs. 40.14) and fewer routes with collisions (86 vs. 108) (Appendix D.1, Table 8).

We believe this comparison provides meaningful empirical support for the contribution of the proposed selection framework.

3. Expanded experimental validation

We also extended our Bench2Drive evaluation to BevAD-M using the same SRTDrive checkpoint trained with VAD, without any additional training or fine-tuning. Under identical evaluation conditions, the Driving Score improved from 88.76 to 90.78, while routes with collisions decreased from 42 to 35.

In addition, we conducted NAVSIM v2 evaluations and Generator component ablations covering progress, normal, and tangential deformations, feature conditioning, and the Rectified Flow decoder (Appendix D.9.1).

We also added a new ablation isolating repair-trajectory supervision (Appendix D.9.2). Under matched 8-epoch training, repair supervision increased the collision-improving candidate generation rate from 41.27% to 55.56%.

We have reported the corresponding results and limitations in the revised manuscript to provide a more complete picture of the framework's behavior across different evaluation settings.

We would especially like to express our gratitude for your comments on novelty. They have helped us articulate more clearly the specific problem SRTDrive is designed to address and distinguish its contribution from the individual techniques on which it builds. We also appreciate your suggestions regarding broader evaluations and component-level analyses, which directly motivated several of the additional experiments.

We sincerely hope that the revised explanations and new experimental evidence help clarify the methodological contribution and practical significance of SRTDrive. If you have the opportunity, we would be truly grateful if you could kindly revisit the revised manuscript and consider these findings in your assessment.

Thank you once again for your valuable time, thoughtful feedback, and constructive suggestions. Your careful review has been especially helpful in strengthening both the clarity and experimental support of our work.

Sincerely,

The Authors

Reviewer_BTB9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a learnable, standalone post-processing module for end-to-end autonomous driving that does not modify the parameters of the underlying autonomous driving model. The module consists of a candidate trajectory generator and a trajectory selector. The candidate trajectory generator takes the trajectory planned by the base model and the scene’s BEV information as inputs and outputs deformations relative to the original trajectory. The multiple candidate trajectories refined by the generator are scored by the selector together with the original trajectory to produce the final planning output. The paper conducts comparative closed-loop experiments on Bench2Drive using VAD, showing that the module improves VAD’s Driving Score from 43.07 to 47.77.

Critical strengths

The paper uses an Adam-based optimizer with privileged ground-truth scene information to obtain repair trajectories, allowing feasible repair trajectories to be generated for each scene according to customizable preferences. These repair trajectories are then used as supervision targets for training the generator. During generator training, the original trajectory, the ground-truth trajectory, and repair trajectories are sampled for supervision. This injects diverse strategies for repairing the original plan into the generator and may mitigate mode collapse in the flow-matching generator.

Critical weaknesses

The comparative experiments in this paper are clearly insufficient. The paper provides closed-loop results only for the relatively early VAD planner. It provides neither additional Bench2Drive closed-loop results on other models nor NAVSIM pseudo-reactive evaluation results, substantially weakening the evidence for the module’s generalizability and effectiveness. Although the generator learns diverse trajectories, Table 4 indicates that, for the vast majority of samples, the selector learns how to select the candidate closest to the ground-truth trajectory, which makes me seriously question whether the diverse strategies learned by the generator ultimately contribute to the final output. Table 4 does not clearly quantify the generator’s actual refinement effectiveness, and I would like to know the proportion of candidate sets with a collision in the original trajectory for which the generator produces at least one candidate with improved safety.

Questions for authors

Among candidate sets in which the original trajectory has a collision, what proportion contain a candidate generated with improved safety? I would very much like to see a more detailed evaluation of this generator’s repair performance. Are there closed-loop results obtained by applying the module to newer, more advanced models?

8 comments
AuthorsOpenReview ↗

Response to Reviewer BTB9 (Part 1/2)

Thank you for your careful review and constructive suggestions. We agree that the original VAD-based closed-loop results alone provided insufficient evidence of generalizability. The key idea of SRTDrive is to use the original trajectory as a common reference for both alternative generation and retain-or-replace decisions, while training the Selector to prioritize avoiding increased collision risk relative to the original. To further evaluate this design, we analyzed how frequently generated alternatives lead to actual collision correction and conducted additional evaluations with another planner and NAVSIM.

1. Do generated alternatives lead to actual collision correction?

When all candidates are collision-free, L2 determines the training target. Therefore, the overall candidate distribution alone is insufficient to assess the effectiveness of collision correction.

Among 5,119 valid nuScenes evaluation inputs, we identified 63 in which the original trajectory had a collision. For these inputs, we analyzed the valid candidates among the seven generated trajectories and their final selection outcomes.

Candidate ConditionInputs with Such CandidatesSelected When Available
Collision improvement relative to the original40/63 (63.5%)25/40 (62.5%)
Collision-free at all evaluation timestamps27/63 (42.9%)20/27 (74.1%)

Collision improvement is defined as no deterioration relative to the original at any of the 1-, 2-, and 3-second evaluation horizons, with an improvement at least at one horizon. Collision-free candidates have no detected collisions at any of the six evaluation timestamps sampled at 0.5-second intervals. The two rows overlap and should not be added together.

Overall, the final output improved the collision metric for 25/63 inputs (39.7%) that originally had collisions. Among these 25 cases, 20 became collision-free at all evaluation timestamps, while 5 showed partial reductions in the collision metric.

The remaining 38 inputs consisted of 23 cases without any improved candidate and 15 cases in which an improved candidate was generated but not selected.

These results quantify how often generated alternatives are actually used for collision correction, while also identifying limitations in both generation and selection. However, the analysis is restricted to the evaluated candidate sets and does not independently establish the contribution of generation diversity itself.

Continued in Part 2/2: Additional evaluation with another planner and NAVSIM.

AuthorsOpenReview ↗

Response to Reviewer BTB9 (Part 2/2)

Continued from Part 1/2.

Update (November 17, 2026): The BevAD-M evaluation results in Section 2 have been corrected due to an unintended checkpoint selection error. Please refer to our follow-up comment, "Correction to BevAD-M Evaluation Results," for the corrected results.

2. Additional evaluation with another planner and NAVSIM

Bench2Drive evaluation with BevAD-M.

We directly applied the SRTDrive weights trained for VAD to BevAD-M [1] and evaluated 220 routes. Both models were kept frozen, and no additional training was performed. We aligned the coordinate ranges and temporal intervals of the BEV features and predicted trajectories.

Model and Result SourceDriving Score ↑Success Rate (%) ↑
BevAD-M — Published Table 388.11 ± 0.9872.73 ± 1.98
BevAD-M + SRTDrive — Additional Evaluation89.4673.18

The additional evaluation was conducted in a single run, and the original BevAD-M was not re-evaluated under an identical environment. Therefore, we do not interpret the numerical difference as a measured performance improvement. Instead, we present this result as an additional closed-loop evaluation demonstrating the application of SRTDrive to another planner without retraining.

NAVSIM v2 evaluation with VADv2.

We also evaluated SRTDrive with VADv2 [2] on 12,146 NAVSIM v2 inputs under the Single-Stage evaluation setting.

The aggregate score changed from 82.167 to 82.166 under the Non-reactive condition and from 81.936 to 81.926 under the Reactive condition. Although NC, DAC, and TTC increased by approximately 0.02–0.03 points, no improvement in the aggregate score was observed.

This evaluation does not include Two-Stage results and should not be interpreted as equivalent to the Bench2Drive closed-loop evaluation, which involves repeated replanning.

We will incorporate the analysis of actual collision corrections and clarify the scope of these additional evaluations in the revised manuscript. We will not generalize these findings into claims of consistent performance improvements across planners or evaluation environments.

Thank you again for helping us clarify the contributions and limitations of SRTDrive.

References

[1] Holtz, D., Hanselmann, N., Doll, S., Cordts, M., and Schiele, B. What Matters for Scalable and Robust Learning in End-to-End Driving Planners? CVPR Findings, 2026. arXiv:2603.15185.

[2] Jiang, B., Chen, S., Gao, H., Liao, B., Zhang, Q., Liu, W., and Wang, X. VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning. ICLR, 2026.

AuthorsOpenReview ↗

Correction to BevAD-M Evaluation Results

We would like to correct the BevAD-M evaluation results reported in our previous response.

We identified that an SRTDrive checkpoint from a different experiment was inadvertently used in the previous evaluation. We apologize for this oversight.

After replacing it with the correct checkpoint, we conducted the evaluation again. The corrected results are as follows:

ModelDriving Score ↑Success Rate (%) ↑
BevAD-M — Baseline88.7672.73
BevAD-M + SRTDrive — Corrected Evaluation90.7875.00

The corrected results demonstrate that SRTDrive improves BevAD-M's Driving Score by 2.02 points and Success Rate by 2.27 percentage points, further supporting its applicability to another planning framework.

These results replace the previously reported BevAD-M + SRTDrive values (DS: 89.46, SR: 73.18). All other results in our previous response remain unchanged.

We apologize for the confusion and thank you for your understanding.

Reviewer_WjHuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SRTDrive, a post-hoc trajectory refinement framework for end-to-end autonomous driving. It keeps a pretrained planner frozen and uses its output as a reference trajectory. A rectified-flow generator produces alternative trajectories by modifying the reference, and a learned selector chooses whether to keep the original trajectory or replace it. The selector is trained with collision-prioritized supervision, which discourages replacements that increase collision risk relative to the reference. Experiments on nuScenes and Bench2Drive show improvements for some planners, particularly VAD.

Critical strengths
  • Improving an existing planner without retraining its weights is meaningful. Keeping the original trajectory as a candidate also allows the method to retain the original plan when no better alternative is available.
  • The reference-based trajectory representation and collision-prioritized selection objective are intuitive.
  • The paper evaluates SRTDrive on Bench2Drive and applies it to seven frozen planners on nuScenes. The closed-loop experiments demonstrate improvements in Driving Score, Route Completion, and the number of routes involving collisions for VAD-Base.
Critical weaknesses
  • The paper introduces a set-contextual selector that compares candidates with each other and with the original trajectory. It also proposes collision-prioritized supervision to train this selector. However, the main closed-loop comparison is against a hand-designed TTC rule. This comparison shows that SRTDrive performs better than the TTC rule, but it does not establish why the proposed selector is better than a simpler learned trajectory scorer. It also does not show which parts of the selector are necessary. For example, the authors could compare the following cases: (1) A standard learned scorer that evaluates candidates independently. (2) A selector without explicit reference-relative features. (3) A selector trained with ordinary collision-aware ranking instead of the proposed lexicographic supervision.

  • The paper uses rectified flow to generate alternatives through progress, normal, and tangential deformations of the reference trajectory. However, it does not compare this approach with simpler methods for generating nearby trajectories. For example, lateral offsets, speed adjustments, or constrained perturbations of the original trajectory could also produce alternative paths without training a flow-based generator.

  • Table 1 evaluates seven frozen planners on nuScenes. However, SRTDrive reduces the collision rate for only three planners and improves L2 error for five. For DMAD, the collision rate increases from 0.147 to 0.161. For SparseDrive and DiffusionDrive, both L2 error and collision rate become worse. Except for VAD, most improvements are very small. Although the authors acknowledge that some planners already perform their own refinement, these results limit the claimed general applicability of the framework. Moreover, a separate SRTDrive module must be trained for each base planner, rather than directly transferring one trained refinement model across planners.

  • The retention and replacement analysis in Section 4.4 shows that more than 99.3% of replacements leave the collision metric unchanged at each prediction horizon. Appendix C.1 reports that 98.178% of the training candidate sets contain only collision-free trajectories. In these cases, the collision-prioritized supervision cannot distinguish candidates based on safety, and selection mainly depends on their displacement error relative to the logged trajectory. Based on these results, how much of the improvement comes from the proposed safety-prioritized supervision, rather than simply selecting trajectories closer to the logged future?

  • In Table 2(a), SRTDrive improves the Driving Score from 43.07 to 47.77 and reduces the number of routes with collisions from 95 to 86. However, Comfort drops from 50.80% to 40.26%. Table 2(b) also shows lower success rates for Overtaking (20.00% to 11.11%) and Give Way (30.00% to 20.00%), although these ability-specific results involve smaller subsets of routes. The paper explains that independently selecting a trajectory at every planning step may cause inconsistent decisions over time. However, this is an important limitation of a refinement method intended to improve an existing planner while preserving its behavior. The authors are suggested to analyze how frequently the selected trajectory changes between consecutive planning steps and whether these changes cause the comfort degradation.

Questions for authors
  • Section 4.4 reports how often the original trajectory is kept or replaced and whether collision metrics change. However, it does not show how many collision cases could potentially be repaired by the available candidates, or how frequently the selector misses a safer candidate.
  • Equation (8) presents the selection loss using cross-entropy and a reference-relative penalty, while Appendix A.6 describes additional loss terms and a separate fine-tuning stage. It would be helpful to clarify in the main paper which objectives are used to train the final reported model and how they relate to Equation (8).
3 comments
AuthorsOpenReview ↗

Response to Reviewer WjHu (Part 1/2)

Thank you for your careful review and constructive suggestions. We acknowledge that the original TTC comparison and aggregate metrics did not sufficiently isolate the contributions of our generation and selection design. The key idea of SRTDrive is to use the original trajectory not only as a reference for generating alternatives, but also as the baseline for learning whether to retain or replace a plan. The original trajectory remains unchanged in the candidate set, while the Selector uses inter-candidate relationships and reference-relative features to make its decision. The training target prioritizes avoiding increased collision risk relative to the original and minimizing collisions, with L2 used to distinguish candidates within these priorities.

1. Does collision-prioritized selection provide benefits beyond GT matching?

When all candidates are collision-free, L2 determines the training target. To isolate the contribution of collision-prioritized supervision, we compared L2-only and collision-prioritized training using identical initial weights, candidate sets, and training budgets. This was a separate experiment in which each Selector was trained for 3 epochs.

Collision-prioritized training maintained the same number of inputs in which an original collision was resolved (10), while reducing replacements that introduced collisions absent from the original trajectory from 6 to 2. Mean L2 increased from 0.6976 to 0.7076 m, whereas the mean collision metric decreased from 0.1519% to 0.1389%. Thus, the observed collision reduction cannot be explained simply by selecting trajectories closer to the ground truth. However, the reduction corresponds to only four inputs and should not be interpreted as a general safety guarantee.

We also conducted additional comparisons to distinguish the proposed Selector from conventional learned scoring approaches.

Selection MethodL2 (m) ↓Mean Collision Metric (%) ↓
Learned Pointwise Scorer0.72180.1433
Without Reference-Relative Features0.72220.1311
Conventional Collision-Aware Ranking0.70330.1514
Full Selector0.6770.125

The Full Selector corresponds to the final model reported in Table 1 of the paper, while the other rows are additional comparisons. The Full Selector achieved lower values on both metrics than all three alternatives.

We further compared the Full Selector with the Pointwise Scorer in Bench2Drive over 220 routes. The Full Selector reported in Table 2 achieved a higher Driving Score (47.77 vs. 40.14) and fewer routes with collisions (86 vs. 108). The original VAD results in the same table were 43.07 and 95, respectively. These results show that simply adding a learned Scorer did not reproduce the improvements obtained with the full selection framework.

2. Do the generated alternatives actually correct collisions in the original trajectory?

We compared the proposed Generator with three simple trajectory modification methods that do not require separate generator training.

Candidate Generation MethodSelected L2 (m) ↓Mean Collision Metric (%) ↓
Lateral Offset0.75220.1476
Speed Adjustment0.78020.1563
Constrained Perturbation0.72250.1530
Deformation-Flow0.6770.125

Deformation-Flow corresponds to our final model, while the other methods were evaluated additionally. The proposed method achieved lower L2 and collision metrics than all three simple alternatives. Importantly, these results measure the final trajectories selected by the Selector, rather than the generated candidates in isolation.

Lateral Offset and Speed Adjustment produced higher L2 errors than the original VAD (0.723 m). Although the original trajectory remains in the candidate set, selection at inference time relies on learned scores rather than ground-truth errors. Consequently, an alternative with higher L2 than the original can still be selected.

To further separate the roles of generation and selection, we analyzed 63 inputs in which the original trajectory had a collision. We defined collision improvement as having no deterioration at any of the 1-, 2-, and 3-second collision evaluation horizons, with a reduction at least at one horizon.

Improved candidates were generated for 40/63 inputs (63.5%), and 25/40 of these inputs (62.5%) resulted in the selection of an improved candidate. Among the 38 inputs without a final improvement, 23 had no improved candidate, while 15 had an improved candidate that the Selector failed to choose.

This analysis shows both the extent to which generated alternatives contribute to actual collision correction and the remaining opportunities missed by the selection stage.

Continued in Part 2/2: Applicability limitations and clarification of the training procedure.

AuthorsOpenReview ↗

Response to Reviewer WjHu (Part 2/2)

Continued from Part 1/2.

3. Applicability limitations and clarification of the training procedure

On nuScenes, collision metrics improved for only 3 of the 7 evaluated planners, with relatively small improvements outside VAD. DMAD showed worse collision metrics, while SparseDrive and DiffusionDrive showed deterioration in both L2 and collision metrics. Furthermore, SRTDrive was trained separately for each planner. We therefore do not present these results as evidence of consistent improvements across planners or transfer without retraining.

We also acknowledge the degradation in Comfort (50.80% to 40.26%) and the reduced success rates for overtaking and yielding.

To further examine planning consistency, we analyzed consecutive plans recorded during SRTDrive driving. After aligning them to the same coordinate system and future timestamps at 1-second intervals, the mean plan change was 1.939 m for the original proposals and 2.058 m for the final selected trajectories. Thus, the measured plan variation increased after selection.

The selected candidate index changed at an average rate of 48.4% per route. However, since candidates are regenerated at each planning step, this is only an auxiliary statistic and does not directly measure actual trajectory switching. These observations alone also do not establish a causal explanation for the Comfort degradation. We will distinguish between preserving the original trajectory as a selection option and guaranteeing the behavior or performance of a replacement.

Finally, we will correct an inconsistency in the description of the training procedure. The final nuScenes Selector was trained for 12 epochs with the Planner and Generator frozen, using the targets defined in Eqs. (6)–(8), Cross-Entropy, and the Reference-Relative Penalty. Collision and L2 priorities are encoded in the training target rather than introduced as additional loss terms.

We will correct the description in Appendix A.6 and the related main text regarding an additional 3-epoch fine-tuning stage and additional loss terms. The 3-epoch comparison discussed above is a separate experiment comparing training objectives, not a fine-tuning stage of the final model. The reported performance results remain unchanged.

Thank you again for helping us clarify both the contributions and limitations of the proposed design.

AuthorsOpenReview ↗

Additional Experimental Evidence Addressing Your Valuable Suggestions

Thank you once again for your thoughtful review and detailed suggestions. We sincerely appreciate the time and care you have devoted to evaluating our work.

Your comments were particularly helpful in encouraging us to examine the contributions of the Selector and collision-prioritized supervision more carefully. We agree that demonstrating improvements over a rule-based TTC selector alone was not sufficient to fully explain the effectiveness of our approach.

Following your suggestions, we have conducted several additional experiments and incorporated the findings into the revised manuscript. We would like to briefly highlight the results most relevant to your concerns.

1. Effectiveness of the proposed Selector

To better distinguish our approach from simpler learned scoring methods, we compared the Full Selector with a Learned Pointwise Scorer, a Selector without explicit reference-relative features, and conventional collision-aware ranking.

The Full Selector achieved the lowest L2 and collision metrics among these configurations on nuScenes (Appendix D.1, Table 7).

Importantly, in the Bench2Drive closed-loop evaluation using the same Generator, the Full Selector achieved a higher Driving Score (47.77 vs. 40.14) and fewer routes with collisions (86 vs. 108) than the Learned Pointwise Scorer (Table 8).

These results provide additional evidence that the proposed selection framework offers benefits beyond simply applying a learned scorer to generated trajectories.

2. Contribution of Collision-Prioritized Supervision

Following your concern about whether the safety improvement could be explained simply by selecting trajectories closer to the logged future, we conducted a controlled comparison between L2-only and collision-prioritized supervision.

Under matched training conditions, collision-prioritized supervision reduced replacements that introduced new collisions from 6 to 2, while maintaining the same number of repaired original collision cases (10).

We believe this comparison helps clarify the specific role of collision-prioritized supervision in the proposed framework (Appendix D.2).

3. Additional Generation Analysis and Cross-Planner Evaluation

We also added comparisons with simpler trajectory modification methods and analyzed how frequently generated alternatives actually improve collisions in the original trajectory.

Furthermore, we conducted an additional Generator ablation isolating repair-trajectory supervision (Appendix D.9.2).

Another important addition is the Bench2Drive evaluation with BevAD-M. Using the same SRTDrive checkpoint trained with VAD, without any additional training or fine-tuning, the Driving Score increased from 88.76 to 90.78, while routes with collisions decreased from 42 to 35.

We have also expanded our analysis of temporal planning consistency and clarified the observed Comfort trade-offs in the revised manuscript.

We would especially like to thank you for raising these questions. They helped us identify which aspects of the framework needed more direct experimental validation, and the resulting comparisons have made the contributions of our work much clearer.

We sincerely hope that these additional results and clarifications help address your concerns. If you have an opportunity to revisit the revised manuscript, we would be very grateful if you could kindly take these new findings into consideration in your assessment.

Thank you once again for your valuable time, careful review, and constructive guidance. We truly appreciate the opportunity to strengthen our work through your feedback.

Sincerely,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions to the Manuscript (Part 1/2)

We would like to express our sincere gratitude to all three reviewers for their careful evaluation of our manuscript, thoughtful comments, and constructive suggestions.

We have carefully considered the concerns and recommendations raised by each reviewer and have revised the manuscript accordingly, incorporating additional experiments, analyses, and clarifications. In particular, we have focused on strengthening the experimental validation, clarifying the key contributions of SRTDrive, and better distinguishing our approach from existing methods.

We respectfully summarize the major revisions addressing each reviewer's comments below.

1. Reviewer WjHu

In response to the reviewer's valuable suggestions regarding the Selector design and training objectives, we conducted the following additional experiments and analyses.

  • Selector Comparisons: We added comparisons with a Learned Pointwise Scorer, a Selector without explicit reference-relative features, and Conventional Collision-Aware Ranking to further evaluate the effectiveness of the proposed Selector design.
  • Generator Comparisons: We conducted additional experiments with simpler trajectory generation methods, including Lateral Offset, Speed Adjustment, and Constrained Perturbation, to examine the effectiveness of the proposed Deformation-Flow approach.
  • Collision Correction Analysis: We quantitatively analyzed how frequently improved candidates are generated and subsequently selected when the original trajectory contains collisions.
  • Additional Analyses and Clarifications: We further examined the effects of collision-prioritized supervision and temporal planning consistency, and clarified the final Selector training procedure and corresponding objectives.

2. Reviewer BTB9

Following the reviewer's helpful suggestions regarding collision correction effectiveness and applicability to additional planners, we extended our experimental evaluation as follows.

  • Collision Correction Effectiveness: We provided a detailed quantitative analysis showing that collision-improving candidates were generated for 40 of the 63 inputs with colliding reference trajectories, and improved candidates were successfully selected in 25 of these 40 cases.
  • Additional Bench2Drive Evaluation: We extended the closed-loop evaluation to BevAD-M, demonstrating improvements in Driving Score from 88.76 to 90.78 and Success Rate from 72.73% to 75.00%, while reducing the number of routes with collisions from 42 to 35.
  • NAVSIM v2 Evaluation: We additionally evaluated SRTDrive with VADv2 on NAVSIM v2 to further examine its applicability across different planning frameworks and evaluation settings.

3. Reviewer SeU6

In response to the reviewer's insightful comments on novelty, component-level contributions, and experimental validation, we made the following revisions.

  • Clarification of Contributions and Novelty: We further clarified the motivation and key contributions of SRTDrive through a direct comparison with ResWorld and additional discussion of related approaches, including CLOVER.
  • Component-Level Ablation Studies: We conducted additional ablation experiments examining the effects of Progress, Normal, and Tangential Deformation, path-local and global scene features, and the Rectified Flow decoder.
  • Extended NAVSIM Evaluation: We expanded the NAVSIM v2 experiments to include component-level evaluations of the Generator and provided additional analyses and qualitative visualizations.
  • Architecture and Presentation Improvements: We revised the architecture figures and related descriptions to more clearly illustrate reference-anchored generation, the candidate set containing the original trajectory, reference-relative selection, and the distinction between training and inference.

Continued in Part 2/2: Revised Manuscript and Appendix.

AuthorsOpenReview ↗

Summary of Revisions to the Manuscript (Part 2/2)

Continued from Part 1/2.

Revised Manuscript and Appendix

To facilitate the review process, revisions and newly added content in the main manuscript have been highlighted in red so that the changes can be easily identified. These highlights will be changed back to black before the final paper submission.

We have also substantially expanded the Appendix to provide additional experimental evidence and detailed analyses. Specifically, we have added a new Section D following the original Section C, with Section D beginning on page 17.

The expanded Appendix includes additional comparative experiments, detailed quantitative analyses, qualitative visualizations, ablation studies, and NAVSIM evaluation results.

We sincerely appreciate the reviewers' thoughtful feedback, which has been invaluable in helping us clarify the contributions of our work, strengthen the experimental evidence, and improve the overall quality of the manuscript.

We would be most grateful if the reviewers could kindly take another look at the revised manuscript and consider the additional experimental results and clarifications.

Once again, we sincerely thank all three reviewers for their valuable time, careful consideration, and constructive feedback.

Update (November 18, 2026): We have revised some wording in the manuscript to avoid potential ambiguity regarding the evaluation setup.

Update (November 18, 2026): We have additionally incorporated the repair-trajectory supervision ablation results on nuScenes and Bench2Drive Dev10 into Appendix D.9.2 of the revised manuscript. Please refer to our follow-up comment for the detailed results and analysis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.