Improving flocking behaviors in street networks with vision

Guillaume Moinard; Matthieu Latapy

REVIEW 5 major objections 7 minor 11 references

Improving flocking behaviors in street networks with vision

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A vision depth of two prevents walker groups from splitting at intersections, and adding depth-six attraction makes groups up to ten times larger on street networks.

desk verdict A transparent, well-scoped extension with a plausible vision mechanism, but the headline result rests on a single untested aggregation rule and no error bars. read the letter →

arxiv 2505.21585 v1 pith:EV2TXNAN submitted 2025-05-27 physics.soc-ph cs.MAcs.SI

Guillaume Moinard , Matthieu Latapy This is my paper

classification physics.soc-phcs.MAcs.SI

keywords streetnetworksflockingvisiondepthalignmentruleattractiongatheringscorerobustnessprotests

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that adding a bounded field of vision to a street-network flocking model turns fragile local rules into robust gathering behavior. In simulations on a discretized Paris street network, walkers who can see two hops of net flux avoid splitting when a group reaches an intersection, because the smaller subgroup still perceives the flux toward the main group and turns back. Walkers who can also see six hops of crowd density actively steer toward distant groups, and the two depths together produce non-empty groups up to ten times larger than the visionless baseline. The paper presents this as a step toward decentralized protocols for protest gathering and other collective urban movement, where individuals have no global information.

What carries the argument

The load-bearing object is the weighted branch. A branch is a maximal simple path of street nodes starting at a walker's current node, and its weight is $w_B(d_n,d_j,t)=N_B(d_n,t)+J_B(d_j,t)$, where $N_B$ averages the number of walkers over the first $d_n$ nodes and $J_B$ averages the net flux over the first $d_j$ links. The walker moves to the first node of the branch with the largest weight, choosing randomly in ties. All vision effects enter through these two averages: depth $d_j=2$ makes the flux term see one link beyond the neighbor, which is enough to detect the main group's movement at an intersection, and depth $d_n=6$ lets the attraction term see several blocks ahead.

What would settle it

Re-run the Paris simulations with distance-decayed branch weights instead of equal averages; if the $d_n=6$, $d_j=2$ gathering score drops to the visionless baseline, the branch-average aggregation is the load-bearing component. Alternatively, observe real pedestrian groups at a two-exit intersection and check whether followers look back at the main group's recent movement; if they do not, the proposed alignment mechanism is not what real walkers use.

Watch

Extended reading notes

Core claim

The paper's central claim is that the vision depth, not just the rule combination, controls whether flocking succeeds on street networks. With the alignment rule at depth $d_j=2$, a group that splits at an intersection does not stay split: walkers in the smaller subgroup still see the positive net flux from the previous move of the larger group and rejoin it, so over long time all walkers flock. With attraction depth $d_n=6$ and alignment depth $d_j=2$, the average size of non-empty groups reaches about ten times the values obtained in the visionless model. The same tactic also recovers from a forced one-step random break-up and ends with larger groups than if no break-up had occurred, which the authors call anti-fragility. The paper attributes these outcomes to the alignment rule preventing divergent splits and the attraction rule pulling distant groups together.

Load-bearing premise

The branch-weighting rule assumes a walker treats every node and link it can see along a branch as equally important and chooses the branch with the largest unweighted average; if nearer information matters more in reality, the reported tenfold improvement may not survive.

Editorial extensions

If this is right

At $d_j=2$, the alignment rule alone suppresses the intersection-split failure mode identified in the earlier model, so the simpler rule is enough for robust long-time flocking.
The tactic $(d_n,d_j)=(6,2)$ gives groups up to ten times larger than the visionless baseline on the tested network.
After a one-step random break-up, this tactic reforms groups larger than before the break-up, not merely back to baseline.
If gathering time is short, attraction alone at high $d_n$ is the better choice; the combined tactic only wins for longer runs.
Increasing $d_j$ beyond 2 degrades the gathering score, because walkers anticipate other groups and steer away before merging.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The branch-weight rule is an unvalidated modeling choice; a distance-decay weighting would likely change which branch wins, so the reported optimum should be tested against other aggregation schemes.
The tenfold claim is demonstrated on one city graph; the dependence of the optimal $(d_n,d_j)$ on network topology is untested, and other cities could shift the optimum.
A natural extension is to make vision depth an individual parameter or let walkers adapt it, which the current homogeneous-depth setup cannot capture.
The anti-fragility result suggests a design principle for distributed gathering protocols: let agents sense one link beyond the immediate decision point, rather than giving them long-range attraction only.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here is what you need to know: this is a reasonable, well-scoped extension of the authors' own flocking model on street networks. Vision is modeled as branch-based perception with separate depths for node counts and link fluxes, and the qualitative finding that alignment vision at depth 2 prevents groups from splitting at intersections is genuinely interesting. The paper is not a breakthrough; it is an incremental step with a promising mechanism.

What is new: the previous model [1] was visionless; this adds a parameterized vision field and shows that a particular combination (attraction depth 6, alignment depth 2) yields larger groups. The branch-weighting rule is precisely defined, the simulation logic is clear, and the implementation is provided. The authors also test robustness to a one-step random breakup and report post-breakup growth.

What is soft, in order of severity. First, all headline numbers come from a single run per parameter cell, with no repetition counts or error bars. The '10 times larger' figure is the best cell of a grid scan; without variance estimates it could be noise, and the paper itself notes that increasing dn above 6 worsens results 'for reasons that are not clear'. Second, the branch scoring uses unweighted arithmetic means (Eqs. 2-3) and no alternative aggregation is tested. The stress-test concern is on target: the dj=2 anti-splitting mechanism relies on a positive flux on a second link offsetting a negative flux on the first; under a max or first-link-only rule the behavior likely disappears. So the central claim is conditional on one modeling choice, and the paper gives no empirical or theoretical justification for unweighted means. Third, the robustness claim ('anti-fragility') rests on one stylized breakup scenario; calling it anti-fragility is a stretch. These are not fatal: the model is transparent, the mechanism is plausible, and the code is a positive step, though without a commit hash or reproduction instructions the exact figures are not independently verifiable.

Who is this for? People modeling pedestrian or protest dynamics on real street networks who want a concrete baseline to compare against. It deserves a serious referee, but one who will ask for repeated runs, variance reporting, and sensitivity analysis on the branch aggregation.

Referee Report

5 major / 7 minor

Summary. The paper extends a previously published street-network flocking model by giving walkers a vision depth, parameterized by dn for the attraction criterion (number of walkers) and dj for the alignment criterion (net flux). A walker evaluates each maximal simple branch starting at its location by the unweighted mean of node counts over the first dn nodes and the unweighted mean of link fluxes over the first dj links, then moves to the first node of the branch with the largest combined weight (Eqs. 1-3). Experiments on a discretized Paris street network (130,276 nodes) compare the gathering score (average non-empty group size at the final step) over a grid of vision depths. The paper reports that alignment with dj=2 prevents group splitting at intersections, that the combination dn=6, dj=2 yields groups up to 10 times larger than previous results, and that such groups show 'anti-fragility' after a simulated break-up.

Significance. The modeling framework is clearly specified, the branch-based vision construction is a natural discrete analogue of neighborhood-based sensing, and the authors provide a reproducibility link with C and Python code. If the headline results were robust, the model would be a useful step toward decentralized flocking on street networks. However, the quantitative claims currently rest on a single run of a grid search and on one arbitrarily chosen aggregation rule; no error bars, multiple trials, or sensitivity analyses are reported, and the proposed mechanism is inferred post hoc from the same runs. The significance of the paper is therefore conditional on additional evidence.

major comments (5)

[§5.2, Fig. 4] The claim that groups are 'up to 10 times larger than previous results' is based on the single best cell (dn=6, dj=2) of an 11x11 heatmap, and the same heatmap is used both to select and to evaluate the tactic. Since no confidence intervals or repeated trials are reported, this value is the maximum over a fitted parameter grid and may reflect noise or overfitting. Please provide means and standard errors over multiple independent runs, and/or a validation protocol that separates parameter selection from evaluation.
[§4.1, Eqs. (1)-(3)] The entire improvement from vision depends on the branch-weighting rule, which uses unweighted arithmetic means of node counts and link fluxes along maximal simple branches. This aggregation choice is neither derived from first principles nor justified empirically, and no alternative aggregation (e.g., max, distance-discounted sum, first-link-only) is tested. In particular, the mechanism described in Section 5.1 for dj=2 — that a walker in a smaller split group perceives a positive net flux because a negative first-link value is averaged with a positive second-link value — is an artifact of the mean aggregation; under a max or first-link-only rule the return behavior would be different or may disappear. Please add a sensitivity analysis over aggregation rules or a formal justification for the unweighted mean.
[§5.1] The explanation that alignment with dj=2 prevents group splitting at intersections is a post-hoc interpretation of the observed group-size curves, not a measured phenomenon. No quantitative evidence about split events, such as the frequency of group splits per intersection encounter or the time a pair of groups remains separated, is provided. Please include direct measurements of splitting and merging events to support the mechanism, and test the prediction that dj=2 specifically reduces splits.
[§5.3, Fig. 6] The 'anti-fragility' claim — that groups turn bigger after a break-up — is supported only by a single random break-up experiment with one trajectory per tactic, and the plotted scores have no error bars. The comparison between the 'without break up' and 'with break up' curves needs repeated simulations and a statistical criterion (e.g., group size after break-up exceeding the no-break-up trajectory by at least two standard errors). Without this, the claim of anti-fragility is not established.
[Abstract, §5] The abstract claims that gathering times are improved, but no quantitative time-to-gather metric is defined or reported. Figures 4 and 5 show final-step group size and group-size evolution, but not a formal measure of gathering time (e.g., time to reach a given fraction of final group size). Please define and report such a metric.

minor comments (7)

[§3.1] The notation G=(V,E) is introduced for both the abstract street network and the discretized network; please use distinct symbols (e.g., G and G_delta) to avoid confusion.
[Def. 4.2] The definition says a branch 'must not be the prefix of any other simple path,' which makes branches maximal simple paths; however, with cycles in the street network, the number of such paths can be exponential. Please state how B(u) is enumerated in the implementation and whether branch sets are pruned.
[Eq. (3)] In Eq. (3), when the branch length |B| is 1, or when dj=0 (both used in Fig. 4), the sum over links is empty and the denominator Δj = min(|B|, dj) - 1 can be zero or negative. Please specify the convention for empty sums and for dj=0.
[§5.2] The statement 'although it is not clear why, increasing dn above 6 also worsen results' is an unexplained empirical observation; since it affects the location of the claimed optimum, the authors should either investigate this non-monotonicity or explicitly list it as a limitation.
[§5.2] The discussion of the 'second column' in the heatmap is hard to follow without numerical values; consider providing the heatmap data in tabular form in an appendix.
[Abstract and §5.1] There are typographical errors: 'guaranties' should be 'guarantees' in the abstract, and 'repealed' should be 'repelled' in Section 5.1.
[§2] Reference [8] by Raymond and Evans is a lattice model, not exactly a line of nodes; the sentence 'this case corresponds to the situation of walkers in a single street' should be qualified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the vision mechanism is a defined model and the reported improvements are simulation outcomes, not fitted predictions.

full rationale

The paper introduces a concrete vision mechanism (branch weighting via Eqs. 1-3) and measures its performance on a discretized Paris street network. The headline improvements, including groups 'up to 10 times larger', are reported outcomes of a parameter sweep in Figure 4, not predictions derived from the model's equations. Selecting the best cell (dn=6, dj=2) from the heatmap and then reporting its score is standard simulation practice; the rule is not defined in terms of that score, so there is no self-definitional reduction. The explanations for dj=2 (e.g., walkers in the smaller split group perceive a positive net flux from averaged link fluxes and return) are post-hoc mechanistic interpretations of observed dynamics. They may be incomplete or overgeneralized, but they do not make the output equivalent to the input by construction. The paper's use of the authors' prior work [1] as a baseline is normal scholarly comparison; no load-bearing claim is justified solely by that citation, and the new results are generated by the simulations reported in this paper. Concerns about the untested branch-averaging aggregation, or about whether the optimal depths generalize to other networks, are robustness and validity questions, not circularity. No specific circular step can be quoted, so the honest finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on two free parameters (vision depths) chosen by grid search, plus modeling choices inherited from the authors' prior paper. The perception model and branch aggregation are ad hoc, with no empirical calibration or alternative tests. These are the main non-derived inputs.

free parameters (3)

attraction vision depth dn = 6 (chosen as optimal from grid search)
Selected as the best cell in Figure 4; performance degrades for dn>6, an effect the authors state they cannot explain.
alignment vision depth dj = 2
Selected as the best alignment depth; dj=1 is the previous model, dj=2 prevents intersection splits, dj>2 causes repulsion.
street discretization length delta = 10 meters
Inherited from prior work [1]; determines graph size (130,276 nodes) and hop-time equivalence.

assumptions (4)

domain assumption The alignment and attraction rules from [1] are effective building blocks for flocking on street networks.
The paper extends these rules without re-testing their adequacy; it relies on the prior model as a baseline and building block.
ad hoc to paper Walkers can perceive the exact number of walkers at nodes and the net flux on links within their vision depth, with no error or cost.
Introduced in Section 4; no empirical basis for this perception model.
ad hoc to paper Branch-based unweighted mean aggregation (Eqs. 2-3) is the correct way to combine multi-step information.
Defined in Section 4.1; no justification or alternative comparison.
domain assumption The Paris OSM graph discretized at 10 m is a representative street network for studying flocking.
Used throughout; the authors acknowledge it is 'a typical instance' (Section 3.1).

invented entities (1)

Vision capacity over branches (vision depth dn, dj)
purpose: Gives walkers information beyond their immediate neighbors, claimed to improve flocking and robustness.
Vision depths are synthetic parameters swept 0-10 and selected by simulation outcome; no real-world vision data or falsifiable prediction outside the model is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving flocking behaviors in street networks with vision." pith.science (2026). https://pith.science/paper/EV2TXNAN

@misc{pith2026250521585,
  author       = {Pith},
  title        = {Pith review of: Improving flocking behaviors in street networks with vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EV2TXNAN}},
  note         = {Machine review of arXiv:2505.21585}
}

read the original abstract

We improve a flocking model on street networks introduced in a previous paper. We expand the field of vision of walkers, making the model more realistic. Under such conditions, we obtain groups of walkers whose gathering times and robustness to break ups are better than previous results. We explain such improvements because the alignment rule with vision guaranties walkers do not split into divergent directions at intersections anymore, and because the attraction rule with vision gathers distant groups. This paves the way to a better understanding of events where walkers have collective decentralized goals, like protests.

Figures

Figures reproduced from arXiv: 2505.21585 by the authors.

**Figure 2.** Walker on node u with d = 2 has access to the information on the sequences of links and nodes highlighted by arrows, as they are the d first nodes and links of the branches in B(u). 4.1 Weighting Branches We know that nu(t) and Jv→u(t) are two commensurable quantities. Indeed, the number of walkers is a quantity that is directly comparable to the net flux, as a group of walkers all moving from u to v at t − 1 produc… view at source ↗

**Figure 3.** Evolution of the last step gathering score for the alignment and attraction rules with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Heatmap of gathering scores for every tactic with combinations of vision depths [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Evolution of gathering score for some relevant tactics with different vision depths. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Plot of robustness experiments, similar to Figure 5. We display the gathering scores for [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

Fast Flocking of Protesters on Street Networks

Moinard, Guillaume and Latapy, Matthieu. Fast Flocking of Protesters on Street Networks. IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 2024

work page 2024
[2]

Flocks, herds and schools: A distributed behavioral model.Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 25–34, 1987

Reynolds, Craig W. Flocks, herds and schools: A distributed behavioral model.Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 25–34, 1987

work page 1987
[3]

Novel type of phase transition in a system of self-driven particles

Vicsek, Tamás and Czirók, András and Ben-Jacob, Eshel and Cohen, Inon and Shochet, Ofer. Novel type of phase transition in a system of self-driven particles. Physical review letters, 75(6):1226, 1995

work page 1995
[4]

Emergentbehaviorinflocks

Cucker, FelipeandSmale, Steve. Emergentbehaviorinflocks. IEEE Transactions on automatic control, 52(5):852–862, 2007

work page 2007
[5]

Consensus and cooperation in networked multi-agent systems.Proceedings of the IEEE, 95(1):215–233, 2007

Olfati-Saber, Reza and Fax, J Alex and Murray, Richard M. Consensus and cooperation in networked multi-agent systems.Proceedings of the IEEE, 95(1):215–233, 2007

work page 2007
[6]

Swarm robotics: Past, present, and future [point of view].Proceedings of the IEEE, 109(7):1152–1165, 2021

Dorigo, Marco and Theraulaz, Guy and Trianni, Vito. Swarm robotics: Past, present, and future [point of view].Proceedings of the IEEE, 109(7):1152–1165, 2021

work page 2021
[7]

Deter- ministic rendezvous in graphs.Algorithmica, 46:69–96, 2006

Dessmark, Anders and Fraigniaud, Pierre and Kowalski, Dariusz R and Pelc, Andrzej. Deter- ministic rendezvous in graphs.Algorithmica, 46:69–96, 2006

work page 2006
[8]

Flocking regimes in a simple lattice model.Physical Review E, 73(3):036112, 2006

Raymond, JR and Evans, MR. Flocking regimes in a simple lattice model.Physical Review E, 73(3):036112, 2006

work page 2006

Show all 11 references

[9]

Collectivebehaviorinanimalgroups: theoreticalmodelsandempiricalstudies

Giardina, Irene. Collectivebehaviorinanimalgroups: theoreticalmodelsandempiricalstudies. HFSP journal, 2(4):205–219, 2008

2008
[10]

OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks.Computers, environment and urban systems, 65:126–139, 2017

Boeing, Geoff. OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks.Computers, environment and urban systems, 65:126–139, 2017

2017
[11]

Random planar graphs and the London street network.The European Physical Journal B, 71:259–271, 2009

Masucci, Adolfo Paolo and Smith, Duncan and Crooks, Andrew and Batty, Michael. Random planar graphs and the London street network.The European Physical Journal B, 71:259–271, 2009. 11

2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fast Flocking of Protesters on Street Networks

Moinard, Guillaume and Latapy, Matthieu. Fast Flocking of Protesters on Street Networks. IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 2024

work page 2024

[2] [2]

Flocks, herds and schools: A distributed behavioral model.Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 25–34, 1987

Reynolds, Craig W. Flocks, herds and schools: A distributed behavioral model.Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 25–34, 1987

work page 1987

[3] [3]

Novel type of phase transition in a system of self-driven particles

Vicsek, Tamás and Czirók, András and Ben-Jacob, Eshel and Cohen, Inon and Shochet, Ofer. Novel type of phase transition in a system of self-driven particles. Physical review letters, 75(6):1226, 1995

work page 1995

[4] [4]

Emergentbehaviorinflocks

Cucker, FelipeandSmale, Steve. Emergentbehaviorinflocks. IEEE Transactions on automatic control, 52(5):852–862, 2007

work page 2007

[5] [5]

Consensus and cooperation in networked multi-agent systems.Proceedings of the IEEE, 95(1):215–233, 2007

Olfati-Saber, Reza and Fax, J Alex and Murray, Richard M. Consensus and cooperation in networked multi-agent systems.Proceedings of the IEEE, 95(1):215–233, 2007

work page 2007

[6] [6]

Swarm robotics: Past, present, and future [point of view].Proceedings of the IEEE, 109(7):1152–1165, 2021

Dorigo, Marco and Theraulaz, Guy and Trianni, Vito. Swarm robotics: Past, present, and future [point of view].Proceedings of the IEEE, 109(7):1152–1165, 2021

work page 2021

[7] [7]

Deter- ministic rendezvous in graphs.Algorithmica, 46:69–96, 2006

Dessmark, Anders and Fraigniaud, Pierre and Kowalski, Dariusz R and Pelc, Andrzej. Deter- ministic rendezvous in graphs.Algorithmica, 46:69–96, 2006

work page 2006

[8] [8]

Flocking regimes in a simple lattice model.Physical Review E, 73(3):036112, 2006

Raymond, JR and Evans, MR. Flocking regimes in a simple lattice model.Physical Review E, 73(3):036112, 2006

work page 2006