acceptodds
Withdrawn from ICLR 2027

RegAD: Rethinking Regressive versus Generative Planning in End-to-End Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

End-to-end autonomous driving has recently embraced generative planners to model multimodal future trajectories, yet their iterative sampling process introduces substantial latency, and their empirical gains are often entangled with stronger architectures and anchor priors. This paper revisits whether heavy generative models represent the only solution for autonomous driving and fully exploits the potential of regressive planners. We identify the performance bottlenecks in regressive planners and optimize them accordingly. To this end, we introduce an anchor-refinement framework and a scorer design that incorporates soft scoring labels and a scoring label switching strategy. Finally, we construct three regression variants (RegAD) corresponding to three representative generative planners. Extensive experiments on NAVSIM and nuPlan demonstrate that RegAD matches or surpasses generative planners under the same model complexity while achieving faster training convergence and lower inference latency. Quantitatively, RegAD achieves 90.2 PDMS on NAVSIM with only one-fifth of GoalFlow’s latency, and 91.34 non-reactive score on nuPlan with one-eighth of FlowPlanner’s latency, both with other modules identical except the planning head. These results suggest that the key design of high-performance multi-modal planning is not necessarily a generative algorithm, but reliable multimodal refinement and candidate scoring. Code will be released upon acceptance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Multimodal Trajectory Planning, Regressive Planning, Generative Models, Trajectory Scoring

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_Uo7kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper questions the implicit assumption that end-to-end autonomous driving must rely on generative planners (e.g., diffusion or flow matching). It argues that, under comparable model complexity, a carefully designed regressive multimodal planner can approach or even surpass generative counterparts while substantially reducing inference latency and accelerating training convergence. Methodologically, the authors start from TransFuser / DiffusionDrive and progressively add anchor refinement and stronger scene interaction (deformable attention, agent/ego interaction, anchor-conditioned modulation, residual refinement). Through a Top‑k oracle analysis of candidate quality, they argue that regression is not inferior in generation upper bound; the main bottleneck is candidate scoring/selection. They then propose soft scoring labels and Scoring Label Switching (SLS: early training uses pre-refinement anchor labels; later training switches to post-refinement trajectory labels). The same regressive recipe is transferred to GoalFlow and FlowPlanner, yielding three RegAD variants. On NAVSIM, RegAD-Transfuser reaches ~87.9 PDMS (on par with one-step DiffusionDrive), and RegAD-GoalFlow reaches ~90.2 PDMS at roughly 1/5 of GoalFlow’s latency. On nuPlan, RegAD-FlowPlanner attains strong non-reactive scores (e.g., 91.34 NR) at about 1/8 of FlowPlanner’s latency, with claimed faster training convergence. Validation mainly includes tables, ablations, latency comparisons, multimodal visualizations, and a closed-loop lane-change case in the appendix.

Critical strengths
  1. A timely, community-relevant control question with multi-baseline empirical support. Amid the rapid default toward generative planners, the paper converts DiffusionDrive / GoalFlow / FlowPlanner into regressive counterparts and reports planning-head latency and training-convergence differences. This kind of “factor isolation” experiment is highly valuable for the community and is the main reason I lean toward weak acceptance.
  2. Diagnostic analysis that pins the bottleneck on the scorer, rather than merely claiming “regression can also be multimodal.” The Top‑k oracle quality curves plus Table 3 (switching is critical) disentangle generation vs. scoring, so soft labels / SLS read as targeted fixes rather than pure engineering glue. This improves both persuasiveness and testability.
Critical weaknesses
  1. Differentiation from existing scoring / vocabulary planning literature is not sharp enough, and the paper insufficiently characterizes regimes where generative methods may still help. Related work mentions Hydra-MDP, WoTE, SparseDriveV2, etc., but the main text does not adequately answer: what is new in RegAD relative to static vocabulary scoring or “scoring is all you need,” and under which interaction/uncertainty regimes multi-step generation remains necessary. The R-mode gap and the failure of transferring SLS to DiffusionDrive hint at such boundaries, but these are not elevated into a systematic conclusion. In addition, generative models can more readily incorporate gradient-based guidance (e.g., classifier guidance) to avoid collisions, which may help explain stronger performance in reactive simulation—an advantage that pure regression schemes generally cannot match in the same way. The authors should explicitly discuss this trade-off in the paper.
  2. Multimodal / safety claims are under-supported by the presented evidence. In §3.3 the authors state that their strategy achieves “lower latency and better multimodal performance.” Later, this is mainly supported by a small number of qualitative visualizations. The amount of visualization is insufficient to substantiate the multimodal claims; quantitative metrics (e.g., pairwise cosine similarity among candidate trajectories) would strengthen the argument. Moreover, in Figure 6, RegAD appears to produce many infeasible trajectories, which conflicts with phrasing such as “Our regression planner generates diverse and safe trajectory candidates.” Given that diverse and safe multimodal trajectories are a central trend in end-to-end autonomous driving, the authors need to provide stronger evidence that RegAD indeed has better multimodal characteristics.
Questions for authors
  1. Please report parameter counts, planning-head FLOPs, and a complete checklist of architectural differences beyond “remove denoising / flow matching and time embeddings” for RegAD-Transfuser vs. DiffusionDrive, RegAD-GoalFlow vs. GoalFlow, and RegAD-FlowPlanner vs. FlowPlanner. If you strip deformable attention / agent interaction / modulation and keep only “MLP residual + your scorer,” how large is the remaining gap to one-step DiffusionDrive?
  2. Implication of SLS failure on DiffusionDrive (clarification; critical): Transferring soft labels + SLS to DiffusionDrive causes a large PDMS drop. Does this imply the central conclusion should be refined to: “under deterministic regression with clean candidate–label alignment, scoring is more critical than generation,” rather than “the generative algorithm itself is unimportant”? Did you try constructing soft labels on denoised / final trajectories (instead of noisy inputs) to restore spatial alignment?
  3. Reactive gap (clarification; critical): What primarily causes the R-mode lag vs. FlowPlanner on nuPlan—interaction modeling, candidate diversity, scorer calibration, closed-loop error accumulation, or FlowPlanner’s ability to use classifier guidance (or similar) to avoid some collisions? Please provide NR/R breakdowns by scenario type (lane change / intersection / close interaction); otherwise the “match or surpass” wording is hard to justify.
  4. Same-backbone comparisons against SparseDriveV2 / Hydra-MDP-style pure scoring systems would better isolate RegAD’s unique contribution.
  5. Report full end-to-end system latency (including perception), not only planning-head milliseconds.
Reviewer_TYJjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work seeks to better understand if generative models (i.e. diffusion planners) should remain the preferred strategy of choice for end-to-end methods citing concerns involving complexity and inference overhead from multi-step denoising steps. The identifies key strategies that can raise regression based model performance while keeping some diversity benefits of generative models. Though there are practical and valuable insights, there are some important points to address

Critical strengths
  • The key contribution of this work focuses on how and if regression based methods can match the performance of diffusion based models, highlighting the pros and cons of diffusion strategies that include inference time due to multi-step denoising processes.
  • The work integrates recent architectural designs as part of ablations to better understand the implications, one main finding a performance improvement is attributed to the scorer component in regression tasks and fundamental to achieving comparable performance to diffusion models
Critical weaknesses
  • The PDMS performance gains between M2-M4 are marginal; multiple runs with different seeds would be beneficial.
  • From Figure 2, we can infer that using regression there are fewer potential candidates. At the same time, the diversity metric is larger than DiffusionDrive. Based on this, it would be good to understand how off the rollouts are from GT when the regression based approach makes a mistake or is scored incorrectly. The paper does not seem to discuss this.
  • Some newer and stronger baselines are missing from the comparison
  • Missing benchmarks with NAVSIM v2 and closed-loop eval
Questions for authors
  • What are the stars on left of Figure 2? The legend is also missing for colors.
  • Connect to weakness 2. How off are RegAD outputs from GT when a mistake is made compared to Diffusion counterparts?
  • Do you have close loop results associated with this work?
Reviewer_kd5nOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies regressive planners for self-driving, identifying its performance bottleneck in trajectory scoring, and closing its gap with generative planners by proposing RegAD and resolving the scoring bottleneck.

Critical strengths

The paper studies the design choice of regressive versus generative planners for self-driving, and demonstrates interesting insight that explains the performance gap between two planners. RegAD is an interesting method that improves planning efficiency. The method can be also applied to other existing generative methods which seems promising.

Critical weaknesses

The problem framing regarding regressive planner versus generative planner has been asked and answered by existing works like SparseDriveV2, Much Ado about Noising.

RegAD performance does not look very competitive. The efficiency gain compared to DiffusionDrive (5.46ms vs 5.61ms) is modest while the PDMS are both 87.90. Some strong baselines seem could also be compared (e.g., ResAD, MeanFuser, DriveDPO, DiffusionDriveV2, SparseDriveV2).

Questions for authors

RegAD Flow Planner seems to have 20 candidates, while the original FlowPlanner uses a single sample. Could the author provide the number of candidates used by FlowPlanner and report the performance with the same number of candidates if their implementation follows the default single-sample settings?

Reviewer_3Dp6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper revisits whether iterative generative planning is necessary for end-to-end autonomous driving. The authors retain the perception modules and overall architectures of three representative generative planners—DiffusionDrive, GoalFlow, and FlowPlanner—while replacing their generative planning heads with anchor-based deterministic refinement and scoring heads.

The proposed method introduces two main changes. First, it employs a multi-anchor refinement head trained with winner-takes-all trajectory supervision. Second, it improves trajectory scoring through soft scoring labels and a training-time scoring-label-switching (SLS) strategy, which switches the scoring targets from pre-refinement anchors to post-refinement candidates halfway through training.

Experiments are conducted on NAVSIM and nuPlan. RegAD-Transfuser achieves performance comparable to the reported one-step DiffusionDrive while maintaining similar planning-head latency. RegAD-GoalFlow achieves a PDMS of 90.2, close to GoalFlow's 90.3, while reducing planning latency from 52.9 ms to 9.7 ms. RegAD-FlowPlanner achieves a non-reactive score of 91.34 on Val14 while reducing planning latency from 79.71 ms to 9.2 ms.

Critical strengths

The clearest empirical result is planning latency at a similar closed-loop score. RegAD-Transfuser ties 1-step DiffusionDrive at 87.9 PDMS (5.5 ms versus 5.6 ms). RegAD-GoalFlow reaches 90.2 PDMS at 9.7 ms against full GoalFlow at 52.9 ms, and RegAD-FlowPlanner reaches 91.34 non-reactive on Val14 at 9.2 ms against FlowPlanner at 79.71 ms.

I would not read those larger gaps as evidence that generative planning is blocked from the vehicle by runtime. The abstract compares against the multi-step versions. Reducing the number of denoising steps while preserving generative accuracy is already an active area of research. Once such step reductions are taken into account, the remaining onboard latency gap may be relatively small. Although latency reduction is a valid systems-level contribution, it does not, by itself, establish that generative planners face a fundamental barrier to real-world deployment.

Critical weaknesses

1、The proposed components provide useful improvements, but their methodological contributions beyond training and supervision strategies could be better clarified. Anchor refinement, soft labels, and scoring-label switching operate on a fixed set of candidate trajectories. On the TransFuser roadmap, the upgraded decoder improves PDMS from 84.0 to 87.3, while soft labels and label switching contribute a further 0.6 gain to 87.9, matching the one-step diffusion head. However, it remains unclear whether these design choices capture essential properties of driving behavior or primarily improve the optimization of a predefined candidate set. 2、The motivation for representing multimodality through an anchor vocabulary and per-frame scoring could be further substantiated. Comparable PDMS scores do not necessarily demonstrate a better representation of multiple plausible futures. Further analysis would be helpful to clarify what the anchors represent, whether the scores reflect beliefs over driving maneuvers or rankings fitted to a single expert trajectory, and how the method handles maneuvers outside the K-means-derived anchor vocabulary. With winner-takes-all supervision on a single ground-truth trajectory, it is also unclear whether the learned modes capture genuinely distinct behaviors or are largely determined by the predefined anchor bank. 3、The temporal consistency of candidate selection also deserves further investigation. At each planning cycle, the method selects the highest-scoring anchor without an explicit constraint on consistency with the previously selected mode. Small changes in the scene could therefore cause abrupt maneuver switches, potentially leading to lateral or longitudinal fluctuations even when individual trajectories are safe. The aggregate metrics in Table 1, particularly the saturated comfort scores, may not fully capture such temporal instability. Evaluating mode consistency and trajectory smoothness across consecutive planning cycles would strengthen the deployment-related claims.

Questions for authors

1、Aside from latency and a matched PDMS, in what sense is anchor refinement plus a scorer a better model of multimodal futures than a generative planner? Mode coverage, score calibration, or behavior when the intended maneuver is missing from the anchor set would change my assessment. 2、Across consecutive planning cycles, how often does the selected anchor identity change, and how large is the jump in the executed waypoint? Comfort is saturated at 100 in Table 1, so it does not answer this. 3、Against generative planners that already use one step or a truncated schedule, the latency gap is about 0.1 ms for DiffusionDrive and about 1.3 ms for GoalFlow. Do the authors still consider runtime the reason to prefer regression on a vehicle, once those step reductions are treated as part of the generative baseline?

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

Related papers

Discussion (0)

Sign in to comment.