acceptodds
Under review as a conference paper at ICLR 2027

Beyond Policy Support: Interaction Constrained Offline Reinforcement Learning for Autonomous Driving

Abstract

Offline reinforcement learning enables reward-driven policy improvement from fixed datasets without requiring online exploration, making it particularly attractive in safety-critical domains. A central challenge, however, is distribution shift: policy optimization may favor actions that are weakly supported by the offline data, rendering value estimates unreliable. Existing approaches primarily control this shift in the policy's own action space. In interactive environments such as autonomous driving, this can be insufficient: a candidate ego trajectory may remain well supported under the marginal behavior distribution while being poorly supported jointly with the surrounding-agent behavior observed in the logged interaction. We refer to this degradation in interaction support as *interaction distribution shift* (IDS), and introduce *Interaction-Constrained Drive Policy* (ICDP), an offline reinforcement learning framework that explicitly controls interaction-level distribution shift. Starting from the joint data distribution over ego and surrounding-agent futures, we show that joint-support degradation decomposes exactly into an ego-support component and a residual interaction-support component. We recover the latter through contrastive density-ratio estimation, isolating interaction compatibility without explicit joint-density modeling, surrounding-agent prediction, or rollouts in reactive simulators or learned world models during policy optimization. Closed-loop evaluations on nuPlan, Interplan and real-world truck experiments show that ICDP suppresses high-value yet interaction-unsupported trajectory selections and improves performance in interaction-critical driving scenarios.

Primary area:
reinforcement learning
Keywords:
Offline Reinforcement Learning, Autonomous Driving, Interaction-Aware Planning, Contrastive Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_h8GEOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. This paper introduces interaction distribution shift (IDS), where an ego trajectory remains supported marginally but becomes incompatible with the surrounding-agent futures recorded in the dataset.
  2. The proposed ICDP framework decomposes joint-support degradation into ego-support and interaction-support components. It estimates the latter through the difference between joint and ego-only contrastive logits, and incorporates this estimate into diffusion-based offline policy optimization.
  3. Evaluation includes reactive and non-reactive nuPlan benchmarks, InterPlan scenarios, and qualitative deployment on a full-scale autonomous truck.
Critical strengths
  1. Clear formulation with a principled estimator. The paper precisely separates marginal ego-support degradation from conditional interaction-support degradation. Using shared negative proposals allows the ego-only score to cancel the proposal-dependent component of the joint score, providing a coherent rationale for the residual estimator.
  2. Encouraging closed-loop results. Under the controlled backbone comparison in Table 2, ICDP improves reactive scores over unconstrained Offline RL from 70.07 to 72.87 on Test14-Hard and from 83.02 to 86.34 on Test14-Random. InterPlan evaluation and qualitative truck deployment broaden the evidence beyond a single benchmark setting.
Critical weaknesses
  1. The central residual-estimation mechanism is not sufficiently isolated. Table 2 compares the complete framework with BC and unconstrained Offline RL, but does not establish the benefit of residual interaction regularization beyond ego-support regularization alone or direct joint-score regularization. Figure 2 also omits a zero-interaction-penalty setting. These controls, together with direct validation of interaction-score behavior, are important for attributing the gains to the paper’s main contribution.
  2. The evaluation leaves uncertainty about robustness and model selection. Table 2 reports variation across ten training checkpoints rather than independent training seeds. This measures temporal variability but cannot establish reproducibility across runs. Moreover, Section 5.2 motivates the final regularization weights using curves evaluated on Test14-Hard and Test14-Random. The authors should clarify whether these choices were independently determined on validation data and report results across independent seeds.
Questions for authors
  1. How does ICDP compare with ego-support-only regularization and direct joint-score regularization under matched training settings? Can the authors directly validate that subtracting the ego score improves interaction-specific discrimination?
  2. Can the authors report independently trained multi-seed results and clarify the validation protocol used to select regularization weights, agent capacity, and checkpoints?
  3. Why does ICDP score only 51.8 on InterPlan’s Nudge Around category, compared with 72.96 for Flow Planner? Does preserving compatibility with logged agent futures discourage useful deviations in these scenarios? A failure analysis would help establish the constraint’s practical boundaries.
4 comments
AuthorsOpenReview ↗

Official Comment by Authors [1/2]

We thank the reviewer for the thoughtful review, and for recognizing the clear formulation and principled estimator, as well as the encouraging closed-loop results under the controlled backbone comparison. We address each point below.


W1 / Q1: Isolating the residual interaction regularization

To isolate the residual-estimation mechanism and attribute the gains to it, the revision adds two components. First, matched ablations of each regularization term, including the zero-interaction-penalty setting now shown in Figure 2 (points 1–2 below). Second, a new validation study of interaction-score behavior (Section 5.3, Appendix E; point 3 below), which tests whether the learned score detects interaction degradation and whether this degradation corresponds to driving failures. For example, across 626 validation scenes, the ego trajectories that our learned interaction score flags as most shifted from the logged interactions (highest quartile of estimated IDS) have more than ten times the safety-violation rate of those flagged as least shifted (16.52% vs. 1.53%).

1. Does residual interaction regularization add benefit beyond ego-support regularization alone? Under matched training, we compare ICDP with ego-only (λI=0\lambda_I = 0) and interaction-only (λE=0\lambda_E = 0) regularization. On the validation split (Table 4), ICDP scores 92.3, compared with 90.6 for ego-only and 91.2 for interaction-only: interaction regularization alone already outperforms ego-support regularization alone, and combining both performs best. The test benchmarks show the same ordering: averaged over Test14-Hard and Test14-Random (NR and R; Figure 2), ICDP reaches 82.64, compared with 81.20 for interaction-only, 80.80 for ego-only, and 80.70 for Base Offline RL. Adding residual interaction regularization to ego-support regularization therefore improves performance by 1.84 points on test and 1.7 on validation, establishing its benefit beyond ego-support regularization alone.

2. How does ICDP compare with direct joint-score regularization under matched training settings? Joint-support degradation decomposes exactly as ΔJ=ΔE+ΔI\Delta_J = \Delta_E + \Delta_I (Eq. 10). Regularizing the joint score directly therefore applies a single weight to both components, which corresponds to the equal-weight setting (λE,λI)=(1,1)(\lambda_E, \lambda_I) = (1, 1). This setting scores 91.4 on validation (Table 4) and 80.70 on test (Figure 2(a)), below ICDP's 92.3 and 82.64 with (0.3,1)(0.3, 1); on test, equal weighting gives no gain over Base Offline RL (80.70). ICDP therefore outperforms direct joint-score regularization, because the residual estimator separates the two components and allows ego shift to be penalized less than interaction shift.

3. Does subtracting the ego score improve interaction-specific discrimination? Interaction-specific discrimination means distinguishing interactions that differ in how the ego and other agents fit together, rather than in how likely the ego trajectory is on its own. We verify the two properties this requires:

  • Subtracting the ego score isolates the interaction component. In a controlled study with known distributions (Appendix E.4), candidates come from a proposal that differs from the data, so the joint score mixes an ego component with the interaction component. After subtracting the ego score, the residual matches the true interaction ratio with R2R^2 of 0.9995 (Gaussian) and 0.991 (nonlinear).
  • The isolated signal discriminates interactions that the ego score cannot. On driving data, we hold the ego trajectory fixed and change only the lead vehicle's future, so any difference between the cases is purely interaction-specific. The ego-only score cannot distinguish them (AUROC 0.50), whereas our learned interaction score (the residual) identifies the lead futures that lead to a safety violation with AUROC 0.99 (Appendix E.3).

Together, these results show that subtracting the ego score yields an interaction-specific signal, and that this signal discriminates degradations in interactions such as safety violations where the ego score alone is uninformative.

AuthorsOpenReview ↗

Official Comment by Authors [2/2]

W2 / Q2: Robustness across independent seeds and model selection

We agree that variation across checkpoints of a single run measures temporal variability rather than reproducibility across runs. The revision therefore retrains BC, Base Offline RL, and ICDP with four independent seeds (point 1) and clarifies that every hyperparameter was selected on held-out validation data, separate from the test benchmarks (point 2).

1. Are the gains reproducible across independently trained seeds? Table 2 now reports four independently trained seeds per method, each evaluated at the same ten checkpoints. The ± is the standard deviation over all 40 scores, so it captures variation across runs as well as across checkpoints. Across these seeds, ICDP improves reactive performance over Base Offline RL from 70.12 to 72.37 on Test14-Hard and from 83.24 to 86.46 on Test14-Random. These gains of +2.25 and +3.22 points are close to the +2.80 and +3.32 reported in the submission, and ICDP also achieves the highest mean in five of six settings and improves over BC in all six. Tables 1 and 7 follow the same protocol. The reactive gains therefore hold across independently trained runs, not only across checkpoints of a single run.

2. Were the regularization weights, agent capacity, and checkpoints determined on validation data? Yes. All choices use a held-out validation split built at the scenario level, which shares no source scenarios with training and no data with the test sets (Appendix B.1, Appendix B.4):

  • Regularization weights: among the four settings compared, (λE,λI)=(0.3,1)(\lambda_E, \lambda_I) = (0.3, 1) achieves the highest NR validation score (92.3; Table 4).
  • Agent capacity: validation scores rise from 89.2 (24 agents) to 92.1 (32) and 92.3 (48), so we use 48 (Table 5).
  • Checkpoints: none are selected; every seed is evaluated at the same ten checkpoints over the final 20,000 updates, and all are reported.

The test-set curves in Figure 2 were not used for selection, and Section 5.2 now states this explicitly. The validation-selected configuration is also the best setting on these test curves, showing that the validation choices transfer to the test benchmarks.


Q3: Nudge Around and the practical boundaries of the interaction constraint

1. Why does ICDP score lower than Flow Planner on Nudge Around? Under the same four-seed protocol (the submitted table averaged checkpoints from a single seed), ICDP obtains 63.80 ± 11.40 on Nudge Around (Table 7), rather than 51.8. The large standard deviation reflects the category's structure: it contains only ten scenarios, and each scenario's score is multiplied by a binary check of whether the ego passes the parked vehicle [2]. A scenario that fails this check scores zero, so one scenario can change the category by up to 10 points. Flow Planner's 72.96, a single reported value, lies within one standard deviation of our mean, and its own ablation reports 64.46 without classifier-free guidance [1, Table 4], comparable to ICDP. ICDP thus retains a lower mean Nudge Around score, while achieving the highest mean overall, High Traffic, and Jaywalk scores on InterPlan.

2. Does preserving compatibility with logged agent futures discourage useful deviations in these scenarios? No. In Nudge Around, the obstacle is a non-moving parked vehicle whose future does not depend on the ego. By the definition of IDS, the interaction penalty therefore does not apply to the obstacle, so nudging around it is not discouraged. Empirically, removing the interaction term (λI=0\lambda_I = 0) gives 60.8, close to ICDP's 63.80 given the category's high variance. Removing the constraint therefore does not close the gap to Flow Planner, indicating that the constraint is not the cause of the lower score.

3. What are the practical boundaries of the constraint? The interaction penalty acts only on agents whose logged futures depend on the ego, and the tolerance ϵI\epsilon_I permits bounded interaction shift. Its boundary lies in beneficial maneuvers that create an interaction with moving agents that is absent from the data, such as merging into a gap that logged traffic never offered. There, the penalty discourages the maneuver unless its interaction shift stays within ϵI\epsilon_I.


We thank the reviewer again for the constructive feedback, which led to clearer ablations and a stronger evaluation protocol. We hope these responses and the revised manuscript address the reviewer's concerns, and we are happy to clarify any further points.


[1] T. Tan et al. Flow matching-based autonomous driving planning with advanced interactive behavior modeling. In Advances in Neural Information Processing Systems (NeurIPS), 2025.
[2] M. Hallgarten et al. Can vehicle motion planning generalize to realistic long-tail scenarios? arXiv preprint arXiv:2404.07569, 2024.

Thank you for the detailed response. The matched ablations and direct validation clarify the contribution of the residual interaction term. The four-seed results and held-out tuning protocol address my concerns about reproducibility and model selection. I will consider raising my score.

We sincerely thank the reviewer for their positive and encouraging feedback. We greatly appreciate the reviewer's thoughtful comments and constructive suggestions, which have helped us further improve the quality and clarity of our paper.

Reviewer_km4ROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The motivation behind this work builds off the limitations of standard offline RL under interactive scenarios. The key part entails how an ego trajectory is in distribution marginally but not in distribution when considering interactions with surrounding agents jointly in a close loop fashion. The key contribution is to reduce this gap in reactive and interactive scenarios.

Critical strengths
  • The formulation appears sound and leverages real logged human trajectories therefore highlighting realistic behaviors
  • Performance appears the strongest in close loop compared to other methods
Critical weaknesses
  • Regarding experimental results in table 2, it is not evident what the authors mean by “unconstrained” offline RL. It is hard to attribute improvements to the designs presented without these details.
  • When comparing performance with other recent models, the BC policy (Table 2) already beats other models without considering ICDP, potentially implying that the network architecture is much stronger. It would make sense to integrate ICDP with other architectural designs to better isolate the benefits.
Questions for authors
  • What features are incorporated in the unconstrained baseline being compared in Table 2? Is it just blocks 1-2 from Figure 1?
  • How are the hyperparameters tuned? Section 5.2 references the test sets as a source of feedback but those details are not entirely clear.
3 comments

We thank the reviewer for the constructive feedback and for recognizing the soundness of our formulation, the use of real logged driving data, and the strong closed-loop performance. We address the reviewer's concerns below.


W1 / Q1: Configuration of the offline RL baseline

"Offline RL," also referred to as "unconstrained Offline RL," denotes our base offline actor–critic model for reward-driven policy improvement from fixed driving data. We have standardized its name as Base Offline RL throughout the revised manuscript.

Base Offline RL comprises blocks 1–2 of Figure 1: the scene encoder, the diffusion actor, and the chunk-level critic ensemble. The diffusion behavior-cloning objective remains active as a behavioral anchor (Appendix B.4). ICDP adds only the contrastive support module in block 3 and its regularization terms.

The controlled comparison in Table 2 therefore separates two effects. BC versus Base Offline RL measures reward-driven policy improvement beyond imitation. Base Offline RL versus ICDP isolates the additional benefit of ICDP's support regularization under the same actor–critic backbone. Section 5.1 and Appendix B.5 now state the modules and objective of each variant explicitly.


W2: Contribution of ICDP beyond the policy backbone

We appreciate the reviewer's emphasis on isolating ICDP's contribution. Our policy backbone is adapted from Diffusion Planner [1], with the prediction horizon and implementation configuration described in Appendix B. The BC baseline scores below Diffusion Planner [1] in four of the six settings and below Flow Planner [2] in five, including all three reactive evaluations in both comparisons.

Table 2 uses the same backbone, prediction horizon, and training data for BC, Base Offline RL, and ICDP. Base Offline RL and ICDP additionally share the actor–critic architecture, driving reward, and NSPG optimization, so their comparison isolates ICDP's support module and regularization (Appendix B.5). This evaluates ICDP on a backbone derived from an existing planner and demonstrates the additional benefit of its support regularization under a fixed actor–critic architecture.

Table 2 now reports results over four independent training seeds, each evaluated at ten checkpoints. ICDP improves reactive performance over Base Offline RL by 2.25 points on Test14-Hard and 3.22 points on Test14-Random, and over BC by 3.96 and 3.65 points. These gains support the benefit of the proposed regularization under a fixed policy backbone.


Q2: Hyperparameter selection and validation protocol

Hyperparameters were selected on a held-out validation split constructed from the nuPlan trainval data: 3,660 offline transitions (Appendix B.1). The split is made at the scenario level: all overlapping windows from a source scenario stay in the same split, so training and validation share no source scenarios. Neither split shares any data with the test sets used for evaluation.

Hyperparameter selection was based on NR closed-loop validation driving scores, obtained by running the planner on the source scenarios of these validation transitions. We compared four regularization settings: (λE,λI)=(0.3,1)(\lambda_E, \lambda_I) = (0.3, 1) and (1,1)(1, 1), plus ego-only and interaction-only variants. We also compared interaction-support contexts of 24, 32, and 48 surrounding agents. Appendix B.4 (Tables 4 and 5) reports the validation scores of all tested settings.

No individual checkpoint was selected. Each seed is evaluated at the same ten checkpoints, spaced 2,000 optimization steps apart over the final 20,000 updates, and Tables 1 and 2 report the mean over all 40 checkpoint evaluations for ICDP.

To assess the diffusion policy's sampled trajectory distribution, we also generated 32 ego trajectories per validation transition and aggregated open-loop collision and TTC statistics. These served as diagnostic checks only and were not used for hyperparameter selection.

Figure 2 reports sensitivity analyses on the test benchmarks; these plots were not used for hyperparameter selection. Section 5.2 and Appendix B.4 now state this explicitly.


We thank the reviewer again for their careful reading and helpful suggestions, which have improved the clarity and rigor of the paper. We hope our responses and the revised manuscript address the reviewer's concerns. We are happy to clarify any further points or provide additional details the reviewer may have.


[1] Y. Zheng et al. Diffusion-based planning for autonomous driving with flexible guidance. In International Conference on Learning Representations (ICLR), 2025.
[2] T. Tan et al. Flow matching-based autonomous driving planning with advanced interactive behavior modeling. In Advances in Neural Information Processing Systems (NeurIPS), 2025.

Thank you for addressing my comments; one question regarding the last note, why were the hyper parameters selected using the validation set instead of the training split?

We thank the reviewer for the follow-up question.

The validation set is a subset of the training data. We split the training data into two disjoint parts:

  1. Training subset: used to optimize the policy, the critic, and our support module.
  2. Held-out tuning subset (3,660 offline transitions): excluded from training and used for hyperparameter selection. In the paper we refer to this subset as the validation set (Appendix B.1).

Hyperparameters were therefore selected using the training data, specifically a part of it that the model did not see during optimization. This follows the standard practice described in Section 5.3 of [1], in which the validation set is constructed from the training data by splitting it into two disjoint subsets, one used to learn the parameters and the other to select hyperparameters. Offline RL algorithms are also known to be sensitive to hyperparameter choices, which makes careful selection particularly important [2]. Neither subset overlaps with the test benchmarks used in the paper (Val14, InterPlan, Test14-Hard, and Test14-Random).

We held this subset out because performance measured on scenarios the policy was trained on can be a biased estimate (an optimistic estimate) of how it generalizes; selecting hyperparameters on the training subset could therefore favor configurations that overfit those scenarios.

We have clarified this split and the terminology in Appendix B.1 of the revised manuscript.


References

[1] I. Goodfellow, Y. Bengio, and A. Courville. Deep Learning, Section 5.3 (Hyperparameters and Validation Sets). MIT Press, 2016.
[2] T. L. Paine, C. Paduraru, A. Michi, C. Gulcehre, K. Zolna, A. Novikov, Z. Wang, and N. de Freitas. Hyperparameter Selection for Offline Reinforcement Learning. arXiv:2007.09055, 2020.
Reviewer_5tGqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Interaction-Constrained Drive Policy (ICDP), an offline reinforcement learning framework for diffusion-based autonomous driving. Its main contribution is to formalize interaction distribution shift (IDS), where a candidate ego trajectory may be marginally supported by logged data but poorly supported when paired with the surrounding-agent futures from the same interaction. The method decomposes joint support degradation into ego-marginal and interaction-specific terms, and estimates the latter by subtracting logits from joint ego-agent and ego-only contrastive classifiers. ICDP incorporates these support estimates into chunk-level critic learning and diffusion-policy optimization. The paper evaluates the method on nuPlan closed-loop benchmarks, InterPlan, and a full-scale autonomous truck. It reports improvements over behavior cloning and unconstrained offline RL, especially in reactive nuPlan evaluation; the real-world evidence is qualitative.

Critical strengths

1、The paper isolates a practically relevant failure mode in interactive offline driving: ego-action support alone does not ensure support for the corresponding multi-agent interaction. The decomposition gives a clear motivation for interaction-aware regularization.

2、The method is evaluated beyond offline metrics, including reactive closed-loop nuPlan results, InterPlan scenarios, and deployment on a full-scale truck. The controlled comparison against BC and unconstrained Offline RL is particularly useful for assessing the added regularization.

Critical weaknesses

1、The central estimator is justified at the population optimum of the contrastive objectives, but the paper provides limited evidence that the learned residual score is calibrated or reliable under finite data and function approximation. The paper should quantify estimator error or stability and connect the estimated IDS to actual joint-support failures.

2、The empirical evidence has important gaps. In the controlled nuPlan comparison, ICDP improves five of six reported settings, but is slightly worse than BC and unconstrained Offline RL on Test14-Hard non-reactive evaluation. The InterPlan table also shows ICDP substantially behind FlowPlanner on NudgeAround (51.8 vs. 72.96), despite leading on other listed categories. The real-world truck section gives qualitative examples without trial counts, quantitative safety/performance outcomes, or uncertainty estimates. These results temper the broad performance claims.

Questions for authors

1、Can you report quantitative validation of the learned IDS estimator, such as held-out contrastive performance and whether higher estimated IDS predicts measurable interaction-support degradation or closed-loop failures?

2、How many real-world trials were conducted in each setting, and what quantitative safety and performance metrics were recorded? Please clarify whether the described trials were used for tuning or only evaluation.

3、What explains ICDP’s lower NudgeAround score on InterPlan relative to FlowPlanner, and how sensitive are the reported gains to random seeds and hyperparameter selection?

2 comments
AuthorsOpenReview ↗

Official Comment by Authors [1/2]

We thank the reviewer for the careful and constructive review, and for recognizing that the paper isolates a practically relevant failure mode in interactive offline driving, that the decomposition clearly motivates interaction-aware regularization, and that the evaluation goes beyond offline metrics, including reactive closed-loop nuPlan, InterPlan, and a full-scale truck. We address each point below. (Response in two parts.)


W1 / Q1: Calibration, reliability, and failure prediction of the learned IDS estimator

We added a dedicated validation study (Section 5.3, Appendix E) that answers the reviewer's questions in four steps, each building on the previous one: we first check that the classifiers are learned reliably, then that their estimate captures interaction, then that it identifies safety violations in driving, and finally we quantify its calibration where the true values are known. For the driving analyses, we train three independent pairs of ego and joint classifiers (different seeds) and evaluate them on held-out validation scenes that are disjoint from training.

1. Are the classifiers reliable on held-out data and stable across seeds? On the held-out validation set (Fig. 3(a), Table 9), both classifiers clearly distinguish recorded driving behavior from candidate trajectories (AUROC 0.83 for ego, 0.84 for joint), with little variation across the three seeds (± 0.02).

Yes: the classifiers generalize to held-out scenes and remain stable across seeds.

2. Does high IDS (degradation) correspond to incompatible interactions? We answer this in two parts (Fig. 3(b), Appendix E.2–E.3): do incompatible motions produce high IDS, and does this signal come from the interaction rather than from the ego trajectory alone?

  • Incompatible motions produce high IDS. The learned interaction score is high for compatible ego and lead motions, such as both braking, and low for incompatible combinations, such as the ego accelerating while the lead brakes, so incompatible motions produce high IDS. Across 1,200 recorded ego–lead pairs, the pairings that actually occurred score clearly higher on average than mismatched pairings.
  • This signal comes from the interaction, not from the ego trajectory alone. We fix the scene and the ego trajectory and change only the lead vehicle's future. An ego-only score necessarily ties across these alternatives (AUROC 0.50), while IDS distinguishes the lead futures that lead to a safety violation from those that do not with AUROC 0.99 (details in Appendix E.3).

Yes: IDS is high for incompatible interactions, and this signal comes from the interaction itself.

3. Does high IDS translate to meaningful driving failures? With recorded traffic fixed (Fig. 3(c), Appendix E.3), we compare alternative ego trajectories within each of 626 validation scenes. The highest IDS quartile has more than ten times the safety-violation rate of the lowest (16.52% vs. 1.53%). On the 124 scenes containing both violating and nonviolating alternatives, IDS ranks violating trajectories above nonviolating ones with AUROC 0.94. Safety labels are excluded from classifier fitting and checkpoint selection.

Yes: IDS distinguishes violating from nonviolating ego trajectories.

4. Is the estimated IDS calibrated where the truth is known? Steps 1–3 show that the estimate is informative on driving data, where its numerical error cannot be measured directly. To quantify this error, we use Gaussian and nonlinear distributions whose interaction density ratios are known exactly. The classifiers learn from finite samples using contrastive labels; exact densities are reserved for evaluation. Across three seeds (Appendix E.4, Table 10), IDS achieves RMSE 0.0544 ± 0.0062 and 0.0892 ± 0.0252, with R2R^2 of 0.9995 and 0.9907.

Yes: the estimated IDS closely matches the true values.

Together, these results demonstrate stable held-out discrimination across classifier seeds, sensitivity to ego–agent interaction, and a strong association between estimated IDS and safety violations in driving replay. The controlled Gaussian and nonlinear experiments additionally demonstrate accurate recovery of known IDS values from finite samples. Further quantitative and qualitative results are provided in Section 5.3 and Appendix E.

AuthorsOpenReview ↗

Official Comment by Authors [2/2]

W2 / Q2 / Q3: Empirical evidence, real-world trials, and sensitivity

1. Is ICDP worse than BC and Base Offline RL on Test14-Hard NR? Across four seeds (Table 2), ICDP outperforms BC on Test14-Hard NR (78.68 vs. 77.18) and improves over BC in all six settings. Against Base Offline RL, Test14-Hard NR is the only exception: 78.68 vs. 78.78. This 0.10-point gap is well within one standard deviation for both methods (± 0.94 and ± 0.84), and ICDP still exceeds all prior planners (best: 76.47; Table 1). In reactive evaluation, where agents respond to the ego, ICDP improves over Base Offline RL by 2.25 and 3.22 points on Test14-Hard and Test14-Random.

In short: ICDP outperforms BC in all six settings; the only exception is a 0.10-point gap to Base Offline RL on Test14-Hard NR, within one standard deviation and reported explicitly in Section 5.1.

2. What explains the lower Nudge Around score relative to Flow Planner? To obtain more reliable estimates, we evaluate ICDP over four seeds, each at ten checkpoints, as in Table 2 (the submitted table averaged checkpoints from a single seed). In the revised Table 7, ICDP obtains 63.80 ± 11.40 on Nudge Around, 62.80 ± 1.80 overall, 70.20 on High Traffic, and 50.70 on Jaywalk. The following observations help interpret the Nudge Around comparison (Appendix C):

  • Nudge Around is small and high-variance. It contains only ten scenarios, and each scenario's score is multiplied by a binary check of whether the ego passes the parked vehicle [3], so one scenario changes the category by up to 10 points.
  • The baselines are single reported values, while ours average 40 checkpoint evaluations. Prior InterPlan results are single values without variance, whereas each ICDP score averages four seeds × ten checkpoints. Given this spread, Flow Planner (72.96) lies within one standard deviation of our mean.
  • Flow Planner's own ablation shows a substantial benefit from classifier-free guidance. Without it, Flow Planner reports 64.46 on Nudge Around and 58.22 overall [2, Table 4], compared with 63.80 and 62.80 for ICDP. ICDP also improves over Diffusion Planner [1] (60.48), whose architecture serves as our policy backbone.

In short: ICDP retains a lower mean Nudge Around score than Flow Planner, while achieving the highest reported overall, High Traffic, and Jaywalk scores among the compared methods.

3. How many real-world trials were run, what was measured, and were they used for tuning? As now reported in Section 5.4 and Appendix D (Table 8), we ran 12 trials on a snow-covered test track: 8 multi-vehicle interactions (overtaking and lead-vehicle following) and 4 standalone curved-road runs. We recorded safety interventions, minimum TTC, and distance travelled. No safety interventions were required in any trial. Interactive trials had a mean per-trial minimum TTC of 2.3 s and a mean distance travelled of 560 m, and standalone trials averaged 440 m. The trials were used only for final evaluation, with no policy training or parameter tuning.

In short: Twelve evaluation-only trials required no safety interventions. We report trial counts and per-setting averages as descriptive evidence of deployment feasibility.

4. How sensitive are the gains to random seeds and hyperparameter selection?

  • Seeds: the controlled nuPlan results and ICDP’s InterPlan results now use four independent training seeds, each evaluated at the same ten checkpoints, with no checkpoint selection (Tables 2 and 7). Standard deviations over these 40 scores range from 0.8 to 2.1 points across the nuPlan settings.
  • Hyperparameters: all hyperparameters were selected on NR validation scores, from a split sharing no source scenarios with training and no data with the test sets (Appendix B.4, Tables 4 and 5). Figure 2 is a test-set sensitivity analysis not used for selection, and the validation-selected configuration, (λE,λI)=(0.3,1)(\lambda_E, \lambda_I) = (0.3, 1) with 48 agents, is also the best setting on the test benchmarks.

In short: on nuPlan, the spread across seeds and checkpoints is at most about 2 points (one standard deviation), and the validation-selected hyperparameters transfer to the test benchmarks.


We thank the reviewer again for the detailed feedback. We hope these responses and the revised manuscript address the reviewer's concerns, and we are happy to clarify any further points.


[1] Y. Zheng et al. Diffusion-based planning for autonomous driving with flexible guidance. In International Conference on Learning Representations (ICLR), 2025.
[2] T. Tan et al. Flow matching-based autonomous driving planning with advanced interactive behavior modeling. In Advances in Neural Information Processing Systems (NeurIPS), 2025.
[3] M. Hallgarten et al. Can vehicle motion planning generalize to realistic long-tail scenarios? arXiv preprint arXiv:2404.07569, 2024.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.