acceptodds
Withdrawn from ICLR 2027

Beyond One Path: Preserving Behavioral Diversity in Flow-Based End-to-End Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Flow-based policies offer a flexible foundation for multimodal trajectory generation in end-to-end driving. However, existing flow-based planning methods can produce nearly identical plans, limiting the useful alternatives available to the planner. We present a framework that addresses behavioral diversity across pretraining, sampling, and post-training. We first examine how flow pretraining shapes trajectory diversity, then construct a fixed noise vocabulary by optimizing coverage in decoded trajectory space across training scenes. We find that fixed noise seeds encode behavioral tendencies that persist across scenarios, while their decoded trajectories adapt to the current scene. Building on this structure, we propose mode-specialized post-training that locally refines each seed’s behavior on scenarios where it produces the highest-reward candidate, rather than optimizing every noise across all scenarios. Cross-scene group-relative advantages guide these updates, while behavior-retention regularization penalizes changes to other useful behaviors.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Trajectory Planning

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 4 reviews, above 27% of the others.
Reviewer_6N2POpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The contribution of this work can be divided into three parts:

  1. analysis of the mode diversity under different imitation training recipes: number of epochs, number of scene tokens, output representation (velocity vs. trajectory) and denoising path (linear vs. GVP).
  2. A shared noise vocabulary selected offline from Gaussian noise to maximize decoded-trajectory coverage and quality across training scenes.
  3. Cross-scene RL: suppress mode collapse in GRPO by grouping the scenes instead of trajectories. Results are validated on three benchmarks demonstrating advanced performance. However, driving performance relies on not only trajectory generation, but also heavily relies on selection, which is underdiscussed in this work.
Critical strengths

The proposed cross-scene RL is novel, and the noise vocabulary is a simple, model-agnostic alternative to handcrafted trajectory anchors. The method shows strong performance on three benchmarks.

Critical weaknesses

Regarding the pretraining recipe, the work shows some empirical results but does not analyze the observations or provide insights into it. Specifically:

  • Is diversity equivalent to quality? The focus of this work is preserving the diversity during pre- and posttraining, while the diversity is not properly measured. The work treats best-of-K performance (line213) as the diversity indicator, but lacks justification.
  • v-pred and x-pred share the same optimal solution and can be converted into each other via Eq. 8. To my understanding, v-pred implicitly upweights low-noise flow timesteps relative to x-pred, by a factor that diverges as from t to 1 under GVP. A deeper analysis of why this leads to higher diversity is needed.

On the cross-scene RL side, the idea is novel but the effect seems to be minor on NAVSIM and HUGSIM.

  • Details of the trajectory selection is missing. How do you train the critic ensemble? How do you weight the predicted subscores to select trajectory? Simply switching the backbone shows more significant improvements (more than 1 PDMS, Tab 6) than the proposed RL method (0.4 PDMS, Tab1).
  • The performance gains seem to be mainly from the better backbone. The SotA method DrivoR uses DINOv2-S as the backbone with a single critic , achieving 93.1 PDMS, while this work achieves 92.55 PDMS with the same backbone with a single critic (Tab. 6), behind DrivoR.
Questions for authors

Please provide details about the critic training and inference

  • how is it trained, what is the supervision targets, is it jointly trained with the trajectory generator, do the critic and generator share the backbone...
  • what's the difference of each critic in the ensemble? training seed, or something else?
  • how are the predicted scores used? How is the trajectory selected?

Can you report results where both the generator and the critic use a backbone comparable to DrivoR (DINOv2-S), so that gains from the proposed method can be separated from gains due to a stronger backbone?

Can you provide the diversity measurement and its correlation to the best-of-K performance?

Reviewer_o3dbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies three challenges in flow-based driving: learning useful multimodal behaviors, covering them with a limited candidate budget, and improving trajectories without losing diversity. It combines a pretraining recipe, a shared noise vocabulary, and local reward-based refinement. The resulting planner achieves strong results on NAVSIM v1 and HUGSIM.

Critical strengths
  1. The reported performance is very impressive: 94.5 PDMS on NAVSIM v1 and 43.4 HD-Score on HUGSIM. The results support the practical value of the overall framework.
  2. The paper examines diversity across the full pipeline, from pretraining to sampling and post-training. This broader perspective is useful because improvements at one stage can reduce the alternatives available at another.
Critical weaknesses
  1. The selected noise vocabulary may be not sufficiently analyzed. What makes a set of noise vectors “good,” and how does its distribution relate to the resulting trajectory distribution? An analysis of this relationship would help explain why the selected noises provide better coverage across scenes.
  2. The evidence for improved diversity relies mainly on planning scores and selected visual examples. A before-and-after visualization of all candidates in the same scenes would help show whether the method produces more distinct, useful alternatives rather than many similar trajectories. Quantitative measures of candidate redundancy and feasible behavioral-mode coverage would further support this claim.
Questions for authors

Could you directly measure behavioral diversity before and after post-training, alongside planning scores? Could you provide a visual analysis of what distinguishes good noise vectors from poor ones?

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

Flow matching has the ability to sample different trajectories from different noise inputs; however, in most cases those trajectories usually concentrate into a few modes. Given a fixed and tight sampling budget, this gives the selector very few samples to select the best options. The authors attempt to tackle this behavior during both pretraining and post-training. In pretraining, the authors opted for velocity prediction, generalized variance-preserving (GVP) flow-time scheduler, compact scene tokens, and reduced training duration to preserve sampling diversity. To promote diverse and high-quality trajectory sampling during post-training, the authors carefully curate a vocabulary pool consisting of K initial noises. Six VGGT-based critics then select the best candidate. The authors then utilize cross-scene GRPO during post-training, grouping different scenes to their corresponding best-performing noise with reward gained over the pretrained model and compared to only scenes from the same noise, promoting diverse yet high-quality trajectories. Main results reveal SOTA performance in three tested benchmarks. Ablation studies show that curated noise outperforms random noise and cross-scene GRPO outperforms standard GRPO.

Critical strengths

The central theme of the paper is promoting trajectory sampling diversity yet preserving quality. The authors have successfully identified key components to improve in both pretraining and post-training and explained thoroughly the purpose of each. Most importantly, the authors have thoroughly backed their claims in extensive ablation studies.

Critical weaknesses

Even though the authors have conducted tests on major benchmarks and achieved SOTA performance, those numbers only indicate the prediction performance against the ground-truth data and nowhere do the authors provide any definitions or metrics for sampling diversity performance. This means the authors’ claim that this paper promotes diversity is unjustified.

Questions for authors

The predicted trajectory performance on said benchmarks is in no doubt SOTA. However, as the title points out, this paper focuses on preserving diversity in flow matching sampling, a key question is what set of metrics constitutes valid measurement for diversity of sampled trajectories. If the said benchmarks do not provide such metrics, it is up to the authors to come with valid definitions for these metrics and provide results on them.

Reviewer_P1gmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to preserve useful trajectory diversity in flow-based autonomous driving under a limited candidate budget. The authors examine how pretraining choices affect generated candidates, then select a shared vocabulary of initial noises and use a learned scorer to choose among the resulting trajectories. During post-training, scenes are assigned to suitable noise identities, and the generator is refined through cross-scene GRPO.

The paper reports competitive results on NAVSIM, HUGSIM, and NavSafe-infinity, with ablations on noise construction, scorer desig

Critical strengths
  1. The paper addresses a useful problem: generating more trajectories does not necessarily provide more useful alternatives. Considering candidate coverage, trajectory quality, and final selection together is relevant to practical planning.

  2. The evaluation includes both non-reactive planning and closed-loop driving. The method achieves 94.5 PDMS on NAVSIM and also performs well on HUGSIM and NavSafe-infinity.

  3. Table 4 reports improvements in oracle and mean candidate scores as well as the selected score, supporting an improvement in the candidate set itself. Table 6 separately examines the scorer's contribution. Reporting these measures separately is helpful.

Critical weaknesses
  1. Using noise inputs to influence generated motions has precedents in [DNO (CVPR 2024)], [OptTrajDiff/OGD (ECCV 2024)], and [UniSteer (2026)]. This is simply context, not a criticism of the method or a request for changes.

  2. The scorer's training distribution may affect the comparison with random noises. I would like to understand which candidate trajectories are used to train the scorer. If training mainly uses the fixed vocabulary, the scorer may be more familiar with those candidates. The selected-score difference from random sampling could then reflect both candidate quality and scorer adaptation.

    Please consider training a scorer on random-noise candidates with the same generator, candidate count, architecture, and training budget. Cross-evaluating both scorers on both candidate sets would help separate these effects. The reported oracle improvement supports better candidate quality, but does not resolve this concern about the selected scores or establish the benefit of keeping the vocabulary fixed.

  3. The behavioral consistency and transfer benefit of fixed noise identities need further analysis. The visualizations do not yet establish whether a given noise retains a consistent behavioral tendency across scenes. Distinct outputs within one scene do not automatically imply that those differences persist elsewhere.

    An inversion-based analysis, as in UniSteer (2026), could help relate noise inputs to motion patterns. This is only a suggestion; my main concerns remain the scorer comparison and the evidence for cross-scene behavioral consistency discussed above. Whether those patterns correspond to stable behavioral tendencies across scenes would still need to be tested. Please report candidate diversity and coverage on unseen scenes before and after RL, including cases where tendencies change or candidates become concentrated. Comparisons should use scenes with comparable decision opportunities. The whole-system cross-domain results are valuable, but do not isolate the vocabulary's transfer benefit.

Questions for authors
  1. How are fixed-vocabulary and random-noise candidates used during scorer training? Do the methods in Table 4 use the same scorer, and does its training data provide comparable coverage of their candidate distributions?

  2. Could the authors provide the following cross-evaluation?

    Scorer training candidatesEvaluate fixed-vocabulary candidatesEvaluate random-noise candidates
    Fixed-vocabulary trajectoriesEvaluateEvaluate
    Random-noise trajectoriesEvaluateEvaluate

    Please match the generator, decoding budget, scorer architecture, supervision, and training budget, and clarify how noise IDs are handled for random candidates. Reporting oracle, mean, and selected scores would help interpret the results.

  3. How consistent are individual noise-conditioned behaviors across comparable scenes? With the generator and scorer frozen, does the vocabulary retain its coverage advantage over freshly sampled noises on a new data domain?

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

Related papers

Discussion (0)

Sign in to comment.