acceptodds
Withdrawn from ICLR 2027

Wider Coverage, Better Plan: Admissible Alternatives for End-to-End VLA Planning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In autonomous driving, most scenes admit several routes to proceed, and a planner, often a Vision-Language-Action (VLA) model today, must recognise these options yet commit to one of them. Recent works enable VLA planners to sample extensive trajectories with generative heads, and on challenging benchmarks, the strongest methods pick one of thousands of candidates with a learned scorer. Can the VLA planner generate one or a small set of trajectories, instead of thousands of alternatives, that already contain high-quality options? We argue that supervised fine-tuning (SFT) on an offline dataset, where a single drive trajectory is sampled per scene, collapses the planner onto one mode. As such, we propose to train the planning head of a VLA planner on admissible alternative plans (AAP), additional plans per scene that a rule-based scorer accepts. Applied to Qwen-Drive-1.0 on the NAVSIM v2 navhard benchmark, AAP improves the single plan from 0.315 to 0.373 EPDMS, a gain that holds when both planners drive at equal speed, and turns near-identical samples into distinct options whose best of 16 rises from 0.359 to 0.614. A second stage of reinforcement learning makes the samples more reliable, and a speed-matched control shows that much of its gain comes from more cautious driving. We advocate evaluating planners by the diversity and coverage of their samples, not only by a single benchmark score. Code will be made publicly available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, End-to-End Planning, Vision-Language-Action Models, Trajectory Generation, Flow Matching, Multimodal Planning, Data Augmentation, Mode Collapse, Reinforcement Learning, GRPO, NAVSIM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tCyqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper observes that generative planning heads in VLA driving models often produce near-copy trajectory samples instead of genuinely diverse candidate plans. It proposes Admissible Alternative Plans (AAP), which generates feasible alternative trajectories using scene geometry and rule-based planning procedures, and filters them according to NAVSIM evaluation criteria. The pipeline first fine-tunes the planning head on these admissible alternatives and then applies GRPO with a diversity reward to further improve sample reliability and coverage. The results show improvements in the single-plan score and a substantial increase in the oracle best-of-16 score. The paper also provides diagnostic analyses separating gains from sample diversity and gains caused by slower, more cautious driving.

Critical strengths

The two-stage AAP training pipeline in Section 5 is practically straightforward. It keeps the VLM frozen and fine-tunes the planning head using admissible alternatives generated from scene geometry and rule-based planning procedures. The approach does not require additional human annotations and can be applied to an existing VLA driving model with limited architectural changes.

Critical weaknesses
  1. The paper is simple with little novelty.
  2. The figures need to be further polished. It's hard to obtain information from these figures.
Questions for authors
  1. How were the lateral offset values, 2.5 s blending interval, speed/jerk configurations, ADE=0.5 m filtering threshold, and AAP target budget selected? Could you provide sensitivity studies showing whether the method remains robust when these values change?
  2. Beyond NAVSIM rule-based filtering, how do you verify that the generated alternatives satisfy vehicle-kinematic and dynamic constraints, such as curvature, steering rate, lateral acceleration, and jerk? Could you report feasibility statistics or failure cases for the generated candidates?
Reviewer_ZESSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Admissible Alternative Plans (AAP), augmenting logged demonstrations with rule-filtered alternative trajectories to fine-tune Qwen-Drive’s planning head. On NAVSIM navhard, AAP improves single-plan EPDMS and oracle best-of-16 coverage. Further experiments examine how reinforcement fine-tuning affects sample diversity, reliability, and driving speed.

Critical strengths
  1. AAP improves single-plan navhard EPDMS from 0.315 to 0.373 without changing the architecture or inference cost, demonstrating the practical value of alternative trajectory supervision.

  2. The analysis goes beyond aggregate scores: sample-quality measurements and slowed-execution controls help distinguish broader trajectory coverage from gains attributable to cautious driving

Critical weaknesses
  1. The paper lacks comparisons with alternative target-generation strategies, such as filtering a fixed trajectory vocabulary or using an existing model. Moreover, one random alternative per scene already matches full AAP, leaving the necessity of multiple alternative targets and the proposed generation procedure unsubstantiated.

  2. The conclusions about single-target supervision causing sample collapse and AAP improving diversity are validated only on Qwen-Drive’s flow-matching planning head. Without experiments on another generative planner, it remains unclear whether these findings generalize beyond this particular model.

  3. The navhard improvement accompanies a navtest decline (87.9 to 86.5, further dropping to 76.7/80.0 after RL). The authors attribute this to more cautious driving but provide no navtest metric breakdown to substantiate it. With non-reactive surrounding agents, the benefits for interactive driving also remain unverified.

  4. The discussion of related work is incomplete, eg Beyond Drive [1], CLOVER[2], Curious-VLA[3]

[1] Beyond Imitation: Learning Safe End-to-End Autonomous Driving from Hard Negatives. ECCV 2026.

[2] CLOVER: Closed-Loop Value Estimation & Ranking for End-to-End Autonomous Driving Planning.

[3] Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models. CVPR 2026.

Questions for authors
  1. How does AAP compare with vocabulary-based or model-generated targets, and why are multiple alternatives needed when one random alternative performs equally well?

  2. Do the findings hold for another generative planner beyond Qwen-Drive?

  3. Which metrics explain the navtest degradation, and do the gains transfer to reactive closed-loop evaluation?

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

The paper fine-tunes the planning head of Qwen-Drive-1.0 on the logged trajectory and additional plans generated from lane geometry, speed profiles and PDM-Closed. Alternatives are filtered using comfort, safety and progress criteria on training scenes. The resulting AAP-SFT model produces a more diverse candidate set, with a large improvement in oracle best-of-16 NAVSIM v2 EPDMS and a smaller improvement in its deterministic zero-noise plan. A second stage uses group-relative, positive-advantage-weighted flow regression. The paper studies how reward optimization, injected noise and slower execution change candidate quality, diversity and coverage.

Critical strengths
  1. The alternative-target intervention has a clear, reproducible construction and substantial evidence for improved coverage. The oracle best-of-16 rises from 0.359 to 0.614 in the main run, and remains between 0.591 and 0.614 across five training seeds. The logged-drive-only fine-tuning control, candidate-family ablations and alternative-mixture sweep help distinguish useful augmentation from merely performing extra optimization. The frozen VLM and unchanged network architecture also limit architectural confounds.

  2. The evaluation openly reports costs and counterexamples to its own motivating story. Oracle selection is explicitly identified as privileged. The paper reports worse average sampled quality after AAP-SFT, a lower navtest score, reduced progress/temporal comfort, and the failure of the second stage to improve the deterministic plan substantially. Slowed-trajectory controls are a valuable attempt to understand benchmark incentives. These observations make the paper informative even without a state-of-the-art deployable score.

Critical weaknesses
  1. The broad direct-planning and speed-independent improvement claims are stronger than the current controlled evidence. The headline 0.373 is the highest of the five reported AAP-SFT seeds; their mean is 0.3506 and one run scores 0.314, versus the base model's 0.315. The paper reports the mean, which is helpful, but the headline and cross-paper superiority claims should reflect this variation. Table 8 also gives only 0.325 when using Qwen-Drive-1.0's published loss, although oracle coverage remains high at 0.634. The matched logged-only control does support an alternative-target benefit under the new recipe; it does not establish that all of the headline gain survives the original loss and representation choices. Finally, Section 6.2 applies the same relative slowdown to each planner's own trajectory, which does not in itself equalize their absolute speed or per-scene travel. Table 3's alleged travel matches are also approximate: 8.8 versus 8.1 m and 10.1 versus 9.5 m. Thus the claimed separation of augmentation or RL effects from caution remains incomplete.

  2. The stated second-stage algorithm is not the cited GRPO objective, limiting the interpretation of that study. Equation (4) minimizes a regression loss weighted by max(A, 0); it omits likelihood ratios, the clipped policy objective, and negative-advantage updates. This is a reasonable form of group-relative reward-weighted self-training, but it is materially different from GRPO/Flow-GRPO. The authors call it reward-weighted flow matching locally while repeatedly labelling the results GRPO. Injecting step noise alone does not resolve the objective difference. Consequently, the results establish behavior of this particular update, not how standard GRPO trades coverage for reliability. Since that interpretation is one of the three stated contributions, the method should either be named and scoped precisely or compared against a faithful GRPO implementation under matched initialization, rewards and compute.

Questions for authors
  1. Can you provide a matched factorial comparison of logged-only versus AAP targets under both the published Qwen loss and your modified loss/representation, with repeated seeds? What exactly differs in the “Qwen-Drive-1.0's published loss” row, including time sampling, normalization and waypoint versus displacement targets?
  2. Do the AAP and second-stage findings persist at matched per-scene travel or along a complete score-versus-progress/travel curve? Please report paired uncertainty over the 225 evaluation groups, rather than treating all 5,912 correlated scenes as independent.
  3. Is Equation (4) the exact update used in every row labelled GRPO, including the step-noise experiment? If so, can the claims be restricted to reward-weighted flow regression, or can an actual likelihood-ratio GRPO baseline be supplied?
  4. In Table 8, is the random single alternative fixed once per scene or resampled during training? Its 0.373 single-plan and 0.617 oracle scores essentially match the full alternative bank. This distinction matters for interpreting how much within-scene target diversity the method requires.

Comments on the paper

1 comment
Area_Chair_8AbDOpenReview ↗

Reminder: Author Response and Reviewer Discussion

Dear Reviewers,

We are halfway through the discussion period. Please review the authors’ responses, address any remaining concerns, and engage in the discussion where appropriate. If your assessment remains unchanged, please briefly acknowledge that you have considered the authors’ responses.

Thank you for your time and contributions.

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

Related papers

Discussion (0)

Sign in to comment.