acceptodds
Under review as a conference paper at ICLR 2027

Collision Avoidance and Shape Fidelity in Flow-Based Swarm Formation

Abstract

Collision avoidance and shape fidelity impose distinct requirements on flow-based swarm formation. We study their relationship using a frozen conditional flow-matching model. On airplane and car, learned second-order corrections with selected damping do not significantly outperform first-order controls. Across training seeds of potential-based corrections, models that produce more collision-free rollouts also place agents farther from the generated reference surface. Two loss interventions alter geometric deviation and collision performance without eliminating close contacts. We then construct a fixed pairwise correction from empirical distance statistics and integrate it with a displacement cap and adaptive sub-steps. Without retraining, it achieves 500 clean rollouts out of 500 across five agent counts up to 512 in each of five ShapeNet categories. Optimal reciprocal collision avoidance, applied within the sampling loop at its released settings, matches these collision counts, but the methods differ in their effects on formation geometry. Ablations show that integration controls affect collision performance and that similar collision counts can accompany substantially different geometric errors.

Primary area:
generative models
Keywords:
flow matching, swarm formation, multi-agent systems

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

This paper studies collision-avoidance methods for generating point cloud, which can be viewed as a trajectory planner for swarms. The original flow-based generative methods is pointwise independent, thus external correction needs to be applied to avoid collision. The authors first analyzed learned correction, and found that second order does not bring in improvements. They then constructed a learning-free pairwise correction, and achieved decent results in experiments. The ablation study further shows the effect of integration control.

Critical strengths

The methodology of this paper makes sense. It first studied the learned correction, and honestly admit its failure cases. The authors then tried a simple learning-free correction methods, and demonstrate that it can achieve good results in experiments. The ablation study also show the important factors which influence the results.

Critical weaknesses

The contribution and novelty of this work is limited. The proposed method is just a simple potential-based pairwise N-body force model penalizing collision. Moreover, the advantage of the proposed method is unclear, as it does not exceed the previous ORCA method in collision avoidance.

Questions for authors

How do the learned correction perform with the same displacement cap and adaptive sub-steps as the fixed correction? In the caption of figure 3, it says "the learned heads are integrated without the fixed correction’s displacement cap and sub-steps". As shown in Table 2, removing the displacement cap drops the fixed correction from 500 to 341 clean rollouts. It is unclear whether the learned corrections fail because of the integration scheme or the model.

1 comment

Thank you for the review.

We agree that the fixed correction is a simple pairwise force. Our contribution is the empirical comparison of correction design and integration, including their different effects on collision avoidance and geometry. We will make that distinction clearer in the revision.

The fixed correction does not consistently outperform ORCA. Lowering ORCA's radius from 0.03 to 0.015 brings its results closer to those of the fixed correction at moderate crowding. Surface distances are within about one reference thickness of each other, and we did not find a statistically significant difference in clean-rollout counts; this does not establish equivalence. Under heavy crowding, mainly airplane at N = 1024 and 2048, ORCA keeps more clean rollouts, while the fixed correction stays closer to the surface. We will report this comparison. One note is that ORCA was observed to be more computationally expensive than the fixed correction, even with the sub-steps. Implementation aside, which was an order of magnitude more expensive, the fixed correction is a pairwise sum, computed with a few tensor operations alongside the flow model, while ORCA solves a small linear program for each agent at every step.

On your question: we have not run the learned corrections with the same cap and sub-steps. In earlier runs without either control, the fixed correction performed better at N = 128 and 256 on airplane: 90 and 83 clean rollouts out of 100, compared with 72 and 31 for the best learned correction. That does not establish whether the gap remains under matched integration. We will add this comparison; until then, we will limit the conclusion to the configurations tested.

Reviewer_YcaNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an empirical comparison of collision avoidance and geometry in flow-based swarm formation, which claimed as the main contributions. Author compare learned first-order and second-order corrections, and propose a fixed correction from empirical distance statistics, and discuss tradeoff between collision performance and formation geometry.

Critical strengths
  1. The paper studies a practical problem motivated by drone formation applications and provides sufficient experiments examining collision avoidance, shape fidelity and effects of correction design and integration choices.
  2. The paper has clear limitation and future work discussion based on the empirical comparison.
Critical weaknesses
  1. The proposed method, the fixed correction does not show clear advantages over the ORCA baseline. Although the fixed correction achieves lower reference-surface distance than ORCA in some settings, it is less reliable than ORCA for some larger or more crowded formations as in Table S10.
  2. Current presentation is still difficult to follow, it will be better if authors can polish the writing and consider move some important supplementary material to appendix or main text instead fully as a separate file.
Questions for authors
  1. Have the authors tested the discussed control of displacement normal to the reference surface to improve the balance between collision avoidance and shape fidelity?
  2. Could the authors better explain and justify why the method keeps the tabulated potential, given the comparable performance of linear-ramp variant?
2 comments

Thank you for the review.

  1. We agree that the fixed correction does not consistently outperform ORCA. Lowering ORCA's radius from 0.03 to 0.015 brings its results closer to those of the fixed correction at moderate crowding. Surface distances are within about one reference thickness of each other, and we did not find a statistically significant difference in clean-rollout counts; this does not establish equivalence. Under heavy crowding, mainly airplane at N = 1024 and 2048, ORCA keeps more clean rollouts, while the fixed correction stays closer to the surface. We will report this comparison and present the fixed correction as a reference for studying correction design and integration.

  2. We will define the fixed correction before the learned corrections, move the learned-correction equations into the main text, and include the supplement as an appendix.

On your questions:

  1. We have not tested control of normal displacement for the fixed correction. In an earlier test on learned corrections (airplane, N ≤ 256), restricting corrections to the surface's tangent plane retained much of the collision benefit for potential-based heads, but little for vector-field heads.

  2. The tabulated potential improved collision counts over the ramp in heavily crowded settings. For example, on rifle at N = 1024, it achieved an average of 45.9 clean rollouts out of 100, compared with 1 for the ramp. At moderate crowding, the ramp is as reliable and closer to the surface. Both use a length scale estimated from the data, and that scale had a large effect on geometry in our ablation (Table 2). We will make this distinction clear and recommend the ramp except under heavy crowding.

Thanks for the reply. I will maintain my rating.

Reviewer_316oOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates sampling-time corrections for pretrained Conditional Flow Matching models to improve collision avoidance while preserving the geometric fidelity of generated point clouds. The authors first analyze learned correction strategies involving first-order and second-order dynamics with vector-field and potential-based formulations, and then propose a fixed pairwise repulsive correction combined with displacement capping and adaptive sub-stepping. Experiments evaluate the effectiveness of these approaches in balancing collision avoidance and shape preservation.

Critical strengths
  1. The paper investigates an important issue about how to satisfy additional physical constraints without substantially degrading the distribution learned by a pretrained generative model. The joint consideration of collision avoidance and shape fidelity is valuable.
  2. The comparison between first-order and second-order corrections, as well as learned and fixed correction strategies, provides useful insights into sampling-time control.The experiments highlight that numerical integration strategies can substantially influence the effectiveness of physical corrections.
  3. The proposed fixed pairwise correction is straightforward to implement and does not require retraining the pretrained Flow Matching model. Its combination with displacement capping and adaptive integration provides a practical solution for improving collision avoidance.
Critical weaknesses
  1. The paper presents the analysis of learned corrections in Sections 3–4 before formally introducing the fixed correction in Section 5, although the latter is already referenced in the earlier analysis. Moreover, the formulations and training details of the learned corrections, including network architectures, objectives, and second-order dynamics, are largely deferred to the supplement. A more self-contained methodological description would improve readability and reproducibility.
  2. The linear-ramp potential achieves the same collision-free rollout count as the proposed empirical log-ratio potential. It is therefore unclear what specific advantage the data-driven potential construction provides over a simple short-range repulsive force.
  3. Although the empirical log-ratio can be interpreted as an energy correction to pair-distance statistics, adding its negative gradient to the pretrained flow field does not necessarily produce the corresponding reweighted distribution. The authors should clarify the theoretical motivation and evaluate how closely the corrected distribution matches the intended target.
  4. The proposed correction should be compared against standard gradient-based collision guidance and projection-based correction under matched integration settings and computational budgets. Additionally, ORCA should be evaluated under comparable collision thresholds and safety margins. Applying the same displacement cap and adaptive sub-stepping to the learned corrections would help isolate the contributions of the correction mechanisms from those of the numerical integration strategy.
  5. Although some geometric metrics are preserved, the substantial changes in 1-NN accuracy suggest that the generated distribution may still be distorted. A more detailed analysis is needed to clarify which distributional properties are preserved and which are affected by the correction.
  6. The paper does not adequately investigate the sensitivity to correction strength ((\lambda=0.01)), cutoff distance ((r_{\mathrm{cut}}=1.5)), displacement cap ((0.15s)), or maximum adaptive sub-steps ((k_{\max}=8)).
Questions for authors

As in weaknesses.

1 comment

Thank you for the detailed review.

  1. Agreed. We will define the fixed correction before the learned corrections and move the learned-correction equations (update rules, parameterizations and training objective) into the main text. We will also include the supplement as an appendix.

  2. The tabulated potential improved collision counts over the ramp in heavily crowded settings. For example, on rifle at N = 1024, it achieved an average of 45.9 clean rollouts out of 100, compared with 1 for the ramp. At moderate crowding, the ramp is as reliable and closer to the surface. Both use a length scale estimated from the data, and that scale had a large effect on geometry in our ablation: holding it fixed raised the surface distance from 0.39 to 1.24 with almost the same collision count (Table 2). We will make this distinction clear and recommend the ramp except under heavy crowding.

  3. We agree that adding −∇U to the flow does not guarantee sampling from the reweighted distribution. We will describe the construction as a heuristic. We have not measured how closely the corrected pair distances match the target.

  4. Lowering ORCA's radius from 0.03 to 0.015 brings its results closer to those of the fixed correction at moderate crowding. Surface distances are within about one reference thickness of each other, and we did not find a statistically significant difference in clean-rollout counts; this does not establish equivalence. Under heavy crowding, mainly airplane at N = 1024 and 2048, ORCA keeps more clean rollouts, while the fixed correction stays closer to the surface. We will report this sweep and note that we did not sweep the fixed correction's own length scale.

    We have not run the learned corrections with the same cap and sub-steps. Earlier runs favored the fixed correction without either control, but they do not establish whether the gap remains under matched integration. We will add this comparison; until then, we will limit the conclusion to the configurations tested. We have not tested endpoint guidance or per-step projection.

  5. MMD and coverage change little, but we agree that this does not establish distribution preservation. The largest 1-NN change is on car at N = 1024, where accuracy rises by 18 percentage points; at N = 2048 it falls by 2.5 points. Changing the closest spacings may contribute to this difference, but we have not isolated that effect.

  6. Table 2 tests removal of the cap and the sub-steps. Earlier runs not included in the paper provide some additional evidence: the potential is zero beyond 1.2 spacings, so any cutoff above that gives the same force; halving the cap reduced clean rollouts at N = 512 from 100 to 93 out of 100; and two or four fixed sub-steps matched the adaptive rule. Results were sensitive to λ before the cap was added, which motivated the cap. We have not run a full parameter sweep under the final settings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.