acceptodds
Under review as a conference paper at ICLR 2027

Learning Smooth Driving Policies via Bézier-Parameterized Flow Matching

Abstract

We propose a kinematically structured flow-matching planner for offline goal-conditioned behaviour cloning of driving data. Rather than directly predicting unconstrained trajectories in the ambient state-action space, the model parameterizes the action component of the learned vector field through low-order Bézier control points for heading and speed, then lifts these actions to form joint trajectories using a differentiable kinematic rollout. This induces a trajectory-level generative model whose samples remain aligned with the kinematics model used for execution, reducing lateral-velocity drift and while maintaining smoothness without relying on post-hoc filtering. The model is trained with a flow-matching objective over full trajectories together with auxiliary supervision on fitted Bézier control points, enabling the use of a compact local motion representation while preserving conditional planning performance under scene and goal contexts.

Primary area:
generative models
Keywords:
Trajectory planning, Autonomous driving, Kinematic curves, Generative modeling

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

The paper introduces a structured flow-matching planner that predicts Bézier control points for heading and speed, then reconstructs positions through differentiable kinematic integration. Tangent matching and optional warm starting encourage continuity, while training combines trajectory-level flow matching with auxiliary losses. Evaluation includes nuPlan closed-loop planning, trajectory smoothness analysis, and an additional adaptation to VaVAM

Critical strengths
  1. Generating positions from predicted heading and speed directly addresses internal trajectory inconsistency. The lateral-velocity and rollout analyses provide evidence that this design improves alignment under the chosen integration model.
  2. The paper includes baseline retraining, reactive and non-reactive evaluations, ablations, and an additional architecture adaptation. Reporting the mixed results helps establish the method’s current limitations.
Critical weaknesses
  1. Limited demonstrated benefit of the incremental parameterization. The base model underperforms the retrained FlowPlanner on all six aggregate closed-loop scores. Its Val14 reactive comfort score also decreases from 86.05 to 81.22, while the VaVAM variants worsen both prediction error and comfort. The central practical benefit therefore remains insufficiently established.
  2. Evaluation protocol and attribution concerns. Appendix D.2 selects guidance weights on Test14-hard, compromising its independence as a test set. Additionally, expert-trajectory-based route filtering is introduced without an equivalently conditioned FlowPlanner comparison, making the source of the final gains unclear.
  3. The method builds on FlowPlanner and combines established curve-based representations with differentiable kinematic rollout. Its novelty primarily lies in this specific integration, whose advantages over simpler structured alternatives are not sufficiently isolated through ablations. Moreover, the mixed closed-loop results and degraded comfort in the VaVAM experiments limit the demonstrated significance of the contribution.
Questions for authors
  1. Can you provide results with guidance weights selected exclusively on validation data and evaluated on an independent test set?
  2. Does the +Route variant use future expert trajectory information during evaluation? How does FlowPlanner perform with the same route conditioning?
  3. Why do improved predicted smoothness distributions not consistently improve closed-loop comfort? Please clarify whether these analyses measure planned or actually executed trajectories.
2 comments

Thank you for your feedback. There is considerable overlap between your comments and that of reviewer tYAM and 3CQG so please refer to our replies there.

  • Please refer to our comments to reviewer tYAM for discussion on comfort.

  • We have corrected the mistake of performing the guidance sweep over Test14-hard and have updated all values using guidance weights chosen based on Val14 as done in the other works, so Test14-hard and Test14 are independent of the selection. With the updated guidance sweeps the base Bezier Planner scores above the re-trained Flow Planner on NR-Val14 (86.89 compared to 86.57) and NR-Test14 (87.11 compared to 85.88), and below it on the remaining four splits.

  • Please see comments to 3CQG about the route filtering. We have edited the text to make the distinction clearer and added additional comparison in the appendix between the filtered routes. We have not yet evaluated Flow Planner under the same route conditioning.

  • Regarding novelty, please see our response to reviewer tYAM.

Thank you for the additional experiments and clarifications. These updates address some of my concerns, but several remain:

  1. It remains unclear whether +Route uses future expert trajectory information during evaluation, and a FlowPlanner comparison with identical route conditioning is still missing.
  2. Closed-loop improvements remain modest and mixed. The explanation for the degraded VaVAM results is speculative and does not explain the relative disadvantage against its baseline.
  3. The added related-work discussion clarifies the contribution, but the distinct benefits of Bézier parameterization over simpler alternatives, and separately from warm starting, remain insufficiently established.

I therefore maintain my rating.

Reviewer_3CQGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a flow-matching planner for offline goal-conditioned behavior cloning in autonomous driving. Instead of directly predicting unconstrained trajectories in the ambient state-action space, the model parameterizes the action component of the learned velocity field using low-order Bézier control points for heading and speed, lifting them to full state-action trajectories via a differentiable kinematic rollout. The model is trained with a trajectory-level flow-matching objective alongside auxiliary supervision on Bézier control-point self-consistency and speed regularization, and is evaluated on the nuPlan benchmark as well as a cross-backbone architecture (VaVAM).

Critical strengths

The paper is clearly structured, integrating low-order Bézier parameterization, heading/speed prediction, inter-segment tangent constraints, and differentiable kinematic integration into a complete trajectory-level flow-matching scheme, supported by both closed-loop nuPlan evaluations (NR and R modes) and cross-backbone experiments on VaVAM.

Critical weaknesses
  1. Contradictory Evidence for the Core "Smooth Driving" Claim and Broad Degradation in Training/Baseline Metrics: In Table 2, the base Bézier Planner achieves a lower Comfort score than the retrained baseline Flow Planner (Re-trained) (NR: 91.59 vs. 93.74; R: 81.22 vs. 86.05), and in the VaVAM experiments (Table 7), all Bézier variants also yield lower Comfort than the baseline (0.050–0.066 vs. 0.084). Figure 4 uses a Savitzky-Golay filtering window different from the standard nuPlan evaluation (Appendix B) without reporting window sensitivity; consequently, Figure 4 is insufficient to explain or support improvements in standard closed-loop Comfort. Furthermore, after modifying the output parameterization, the retrained baseline (Flow Planner (Re-trained)) drops substantially in Table 1 compared to the original model, yet the base Bézier Planner still underperforms this weakened baseline across all 6 metrics in Table 1. In the Table 4 ablation study, adding the self-consistency and speed regularization losses under No warm actually degrades the closed-loop score (and the two baseline rows in Table 4 report contradictory scores).

  2. Questionable Kinematic Executability Motivation and Evaluation Protocol Compliance Issues: First, Section 5.1 directly equates the steering angle δ\delta with the heading angle θ\theta (δ=θ\delta=\theta), and Eq. (6) only integrates position using the predicted speed and heading, lacking the wheelbase-dependent heading evolution and physical steering constraints of a bicycle model. This design only guarantees internal geometric consistency (vlat≈0v_{lat}\approx 0) and is insufficient to prove reduced execution drift during actual vehicle tracking. Second, the evaluation protocol undermines the credibility of the conclusions: Appendix D.2 explicitly sweeps and selects guidance weights on the Test14-hard test split (Tables 5 and 6); meanwhile, +Route filters the test route corridor within 3m based on the ground-truth expert trajectory (Section 6, Appendix C.4), leaving it unclarified whether future expert information is leaked during closed-loop testing and whether baselines receive the exact same conditions. Regarding novelty, subsequent versions of MPD already generate smooth trajectories via low-dimensional B-spline representations; the incremental contribution of this paper rests on the specific combination of heading/speed parameterization with trajectory-space flow matching and rollout, whose practical advantage is currently insufficiently demonstrated.

Questions for authors
  1. Comfort Degradation and Ablation Contradictions: Why does the base Bézier parameterization lower Comfort in Table 2 (91.59 / 81.22 vs. 93.74 / 86.05) and across all VaVAM variants in Table 7 (0.050–0.066 vs. 0.084)? Also, why do the auxiliary losses degrade No warm scores in Table 4, and why do the two Bézier Planner baseline rows in Table 4 have conflicting values?
  2. Evaluation Protocol and +Route Conditioning: During closed-loop evaluation, does +Route rely on future ground-truth expert trajectory states to filter the route corridor, and were the baselines evaluated under the exact same filtered route conditions? What are the test scores on Test14-hard and Test14 when guidance weights are selected strictly on a validation set?
  3. Kinematic Executability vs. Internal Alignment: Given that Eq. (6) sets δ=θ\delta=\theta without wheelbase heading evolution or steering limits, how does this formulation guarantee physical vehicle executability or reduce tracking drift under the actual closed-loop bicycle controller?
2 comments

We thank the reviewer for their detailed comments. All values below refer to the updated tables, after re-running the guidance sweeps on Val14.

  • With respect to weaknesses and questions on comfort please see our responses to reviewer tYAM and the following response. The nuPlan comfort score is computed closed-loop on the executed ego states at 10hz and the LQR tracker computes each acceleration command from the planned speed at a 1 s lookahead, so the score depends on how much consecutive plans differ rather than on the smoothness of a single plan. Consistent with this, +Warm, which only constrains the initial heading and speed of consecutive plans, raises comfort from 92.49 to 94.19 (NR) and from 84.44 to 86.14 (R), above the re-trained Flow Planner (93.56 / 85.24). For the VaVAM experiments all learned models, including the baseline, score far below log replay (0.050-0.084 vs 0.729); we speculate this is due to the comparatively slow 2hz sampling rate of the data used for training and evaluation leading to increased jerk and state this in Appendix E.

  • We have corrected the mistake of performing the cfg sweep over test14 and have updated all values using cfg weights chosen based on val14 as done in the other works. With the updated guidance sweeps the base Bezier Planner scores above the re-trained Flow Planner on NR-Val14 (86.89 vs 86.57) and NR-Test14 (87.11 vs 85.88), and below it on the remaining four splits.

  • With respect to the training objective ablation in Table 4, without warm start the auxiliary losses lower the score slightly (-0.62 and -0.74), whereas with warm start they improve it (+1.12 and +0.67); since the final model is sampled with warm start we kept both losses. It should be noted these are separate runs so the results do not correspond to any of the results in the main tables.

  • With regards to the kinematic model, there was a mistake in the text that did not align with our implementation of the bicycle model. We have corrected the main text to reflect this. The used formulation is correct and equivalent to those we cite with steering angle defined implicitly through the path curvature and rear axle to centre of gravity lr.

  • Regarding +Route we have added additional text in the appendix describing this and have edited the main text for clarity. We have also provided, in the appendix, a set of randomly selected comparisons between the two route conditioning methods. Based on the renders the filtering appears to offer increased localization for select low speed trajectories but can also add confounding lane change information for some samples.

Thank you for the rebuttal. It addresses some of my concerns, but it is still unclear how much of the reported improvement comes from the Bézier parameterization itself. I will keep my score.

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

In the context of autonomous driving, the paper proposes a Bezier parameterized flow matching policy formulation that outputs smoother trajectories than unconstrained naïve flow-matching policies.

Critical strengths
  1. The problem is well motivated
  2. The method appears to be sound
  3. The abstract and introduction are well written and easy to follow.
Critical weaknesses
  1. Overall results are not too convincing: the superiority of the proposed method is not well supported, and the experimental analysis is not sufficiently thorough. a. The main claimed advantage of dynamically feasible trajectories seems to be improved comfort (L375-). This is supported by results in Fig 4, however, the exact opposite trend is shown in Table 2 where comfort scores for the proposed method are lowest among all alternatives. The results in Table 2 are not discussed anywhere in the text. b. Overall driving score is lower or on-par with alternatives. This itself might not be an issue but the reasons are not well explained. c. Given that the overall driving scores are not improving SOTA and comfort scores are particularly poor, I would expect a much more thorough experimental analysis.

  2. Novelty might be narrower than suggested by the paper, given the following prior works from the robotics literature: a. Movement Primitive Diffusion (Scheikl et al., 2024), similarly performs sequence-space diffusion using a structured primitive decoder with initial-condition constraints, although it does not include the proposed vehicle-kinematic rollout. b. Motion Planning Diffusion (Carvalho et al., 2025) learns compact, smooth spline-based trajectories, but diffuses directly over spline coefficients.

Questions for authors

In real world systems trajectory predictions are tracked by a downstream controller. While intuitively, having a smoother trajectory plan can be expected to lead to better closed-loop trajectory after controller, it is unclear if this actually happens in practice, and how much of the benefit of smooth predictions get eliminated in the presence of the controller that itself acts as a form of natural smoother. It would be interesting to discuss and evaluate this issue – it might be also the reason for the lower comfort scores despite smoother trajectories?

2 comments

Thank you for the critical feedback.

  • In an attempt to address the reviewers first point, we view the primary aim of this work as a demonstration of how to effectively train a DiT architecture, like that used in Flow Planner, in an end-to-end manner using a differentiable simulator and a reduced order kinematics-models in order to produce constrained kinematically feasible trajectories. Then under this setting, is there a meaningful performance gain to be had given better kinematic alignment from predictions in the form of reduced drift from expert examples. The comparison over nuPlan for Flow Planner, our retrained version, and our model shows, for our best variant (+Warm & Route), performance lies -4.9 to +1.8 points across the evaluation splits.

  • With regards to comfort, Fig.4 and Table.2 measure different quantities. We have edited the text to make this more clear and added additional detail in the appendix. The nuPlan comfort score in Table.2 measures the closed-loop comfort of an auto-regressively replanned ego trajectory tracked using nuPlans LQR tracker with thresholds. While Fig.4 computes open-loop predicted trajectories. Within Table 2, the base model has the lowest NR comfort (92.49), whereas +Warm and +Warm & Route reach 94.19 NR, above the re-trained Flow Planner (93.56) and PlanTF (93.67).

  • We were not aware of either of the papers the reviewer suggested. We have edited the related work section to include these and added a paragraph to highlight the distinctions between these works and ours.

  • With regards to the reviewers final question, this served as our initial motivation for this work. In nuPlan the planner is re-queried every 0.1 s and the LQR tracker computes each acceleration command from the planned speed at a 1 s lookahead, so the executed jerk depends on how much consecutive plans differ rather than on the smoothness of any single plan, which is what Fig. 4 measures. Consistent with this, the +Warm variants, which only constrain the initial heading and speed of consecutive plans, raise comfort from 92.49 to 94.19 (NR) and from 84.44 to 86.14 (R) in Table 2 without changing the model.

Thank you for the rebuttal, i appreciate the effort. The paper would improve with the new evidence, i think this is an interesting direction, but overall the paper remains under the bar for acceptance. Keeping my score

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.