clockwise — does counterclockwise crowd motion emerge in JuPedSim?

A model-based test, using JuPedSim, of the claim by Echeverría-Huarte et al. (2026) that the spontaneous counterclockwise (CCW) rotation of a freely roaming crowd comes from an individual locomotor bias, not from pedestrian interactions.

Status: design complete, implementation pending. Results below are placeholders to be filled in as the study runs. See docs/design.md for the full design and materials/ for the source paper and data.

The phenomenon

When people roam freely inside a confined circular arena, the crowd drifts counterclockwise. The paper measures this with a polarization parameter M: each person's velocity projected onto the counterclockwise (azimuthal) direction around the arena centre, averaged over the crowd. M̄ > 0 is CCW; the experiments find a robust M̄ ≈ 0.2. The authors argue this is not an emergent effect of people avoiding each other — it persists even for a person walking alone — but a slight per-person left-turn bias.

The question

JuPedSim's operational models are pure interaction models: agents avoid collisions but have no built-in turning bias. That makes the simulator a clean testbed:

• Control — agents roam with collision avoidance and a symmetric wall response (turn toward the centre). If the paper is right, no CCW rotation should appear (M̄ ≈ 0).
• Biased — give agents the paper's proposed individual bias: turn left when facing the wall. If the mechanism is right, a CCW rotation should appear (M̄ > 0, near the experimental ≈ 0.2 for a calibrated turn strength).

The model and our decisions

• Arena: a walkable disk of radius 5 m (matching the Spanish experiment), no internal obstacles.

• Roaming: each agent follows a heading that random-walks (unbiased wander). Near the rim it turns away from the wall — symmetrically toward the centre in the control, or left (CCW) in the biased condition. Agents are moved by JuPedSim direct steering toward a point just ahead of them (clamped to stay inside the disk). The behaviour layer is ours; JuPedSim handles only collisions — so the model is "a model on top of a model". One agent, with the target we feed it each step (python scripts/steering_demo.py):

  

• Interactions: collision avoidance is handled by the Anticipation Velocity Model (AVM), JuPedSim's richest lateral-avoidance model — so "no CCW from symmetric avoidance" is a strong result.

• Where the bias lives: we try the individual bias in two places. (a) Wall response — turn left only when facing a wall; a calibration spike confirmed a symmetric wall-turn gives M̄ ≈ 0 while a leftward one gives a strong M̄ > 0, so this is a single knob (biased_fraction) cleanly separated from the collective (AVM) avoidance. (b) Free-space veer — a constant leftward curvature applied every step (free_curvature), the paper's "intrinsic" reading. We compare both against the experiment's spatial field below; they fail in different, informative ways.

• Metric: M(t) exactly as in the paper (azimuthal projection of each agent's velocity, averaged), with M̄ the time average after a warm-up.

• Conditions: control (symmetric) vs biased (turn-left-at-wall), at crowd sizes N ∈ {16, 24, 32}, several seeds each.

• One calibrated parameter: the leftward wall-turn strength, tuned so the biased case lands near M̄ ≈ 0.2; we report its value. The headline result is the qualitative contrast, not a fitted number.

Scope (honest): this reproduces the confined-arena CCW via the wall-turn mechanism. The paper also finds CCW without boundaries and for lone walkers, which a wall-turn model does not explain — so we test the paper's confined/wall hypothesis specifically, not its full claim. Full rationale and risks are in docs/design.md.

Results

5 m arena, N ∈ {16, 24, 32}, 10 seeds per point. (Reproduce with python -m clockwise --fractions 0.0 0.25 0.45 0.7 1.0 --sizes 16 24 32 --seeds 10 --out docs/results.)

1. Symmetric avoidance gives no rotation; a left-turn-at-wall bias does. With nobody biased the crowd does not rotate (M̄ ≈ 0 at every size); as the share of left-turners grows, M̄ rises monotonically. A share of ~35 % reproduces the experiment's M̄ ≈ 0.2; if everyone turns left, the rotation is much stronger (M̄ ≈ 0.57).

share of left-turners	mean M̄ (N=16)	(N=24)	(N=32)
0 % (control)	−0.01	0.00	0.00
25 %	0.16	0.17	0.16
45 %	0.28	0.27	0.26
70 %	0.46	0.37	0.40
100 %	0.61	0.58	0.53

The flat control is not specific to one model. The avoidance is symmetric in every JuPedSim operational model, so the no-bias control should be flat in all of them — not just AVM. Running the control through SocialForceModel, WarpDriver, CollisionFreeSpeedModel and AnticipationVelocityModel, each lands within ±0.03 of zero, far below the experimental M̄ = +0.185. The rotation is not a property of any particular collision model; it has to be put in.

The same point as a video — one real experimental run beside the four model controls. The real crowd circulates counterclockwise; the bare models just mill:

And the converse: give every model the same bias (30 % left-turners) and all four pick up a counterclockwise drift — the rotation is put in the same way regardless of the collision model:

Reproduce with python scripts/compare_models.py and python scripts/collage_video.py.

2. The polarization distribution shifts CCW, as in the paper's Fig 2. The control M(t) is centred on zero; with 45 % left-turners it shifts to M̄ ≈ +0.23.

3. The rotation is visible. Three crowds side by side as the share of left-turners increases — the control mills without a net sense; with more left-turners the crowd circulates counterclockwise. Each agent is coloured by its own rotation (blue = CCW, red = CW) and the collective polarization M ticks in the corner:

Regenerate with python scripts/make_rotation_media.py. Individual cases (full-quality MP4s in docs/results/):

0% — control (M̄ ≈ 0)	45% — ≈ paper (M̄ ≈ 0.28)	100% (M̄ ≈ 0.62)

What this shows. JuPedSim's collision avoidance is symmetric and produces no preferred rotation; adding the paper's proposed individual bias — turning left when facing a wall — is sufficient to make a confined crowd rotate counterclockwise, and the magnitude is set by how common that bias is in the population. This matches the paper's confined-arena result. It does not address the paper's boundary-free and lone-walker findings (see Scope above) — a wall-turn mechanism cannot.

Validation against the experimental data

The authors' trajectory files include a per-agent polarization column (Pol). Recomputing it from their (X, Y, VX, VY) with our metric matches exactly (max difference 0.0000, correlation 1.0 across every Spanish trial), so our M is identical to theirs. Pooled over all Spanish trials the experimental M̄ = +0.185. Choosing the share of left-turners so the simulated M̄ matches (≈ 30 %), the full M distribution — not just the mean — closely coincides with the experiment:

Reproduce with python scripts/validate_against_data.py (needs materials/ExperimentalData/ unzipped).

Where the rotation lives — and two ways a minimal model fails

Matching the mean M̄ is not the same as matching the mechanism. Their analysis code lets us go further and ask where in the arena the rotation happens, by mapping the mean local m over space (the paper's Fig 3 idea). We compare the experiment with two minimal models:

• wall-turn — a share of agents turn left only when facing the wall (the model calibrated above; M̄ matches the experiment by construction).
• intrinsic veer — every agent has a faithful, constant left veer applied at every step (the "individual locomotor bias" the paper proposes; ref. 37, walking straight into circles), with a symmetric wall response. We report what this actually does, not a tuned match.

(The colour scale is clipped at ±0.4 for visibility; the wall-turn rim actually reaches m ≈ 0.98, far past the clip — see the profile below.)

Mean local m per ring (distance r from the centre, in metres):

r (m)	experiment	wall-turn	intrinsic veer
0.25	+0.05	+0.01	+0.06
0.75	+0.02	−0.00	+0.05
1.25	+0.06	−0.01	+0.03
1.75	+0.10	+0.02	+0.04
2.25	+0.11	+0.06	+0.01
2.75	+0.20	+0.04	−0.10
3.25	+0.21	+0.09	−0.21
3.75	+0.33	+0.24	−0.12
4.25	+0.21	+0.62	−0.10
4.75	+0.15	+0.98	n/a
M̄	+0.185	+0.203	−0.097

(The intrinsic model's outermost ring is empty — the symmetric wall response keeps agents off the very rim.)

The experiment shows a coherent counterclockwise rotation that fills the disk: positive at every radius, building from a faint core to a peak in the outer-middle (r ≈ 3.75 m), then easing at the very wall. Neither minimal model reproduces this, and they fail in different ways:

• wall-turn gets the amount of rotation right but puts it in the wrong place — a thin edge spike at the rim, with a near-still interior. The bias only fires at the wall, so only agents at the wall rotate.
• intrinsic veer concentrates its rotation in the outer-middle (around r ≈ 3.25 m) rather than a thin rim spike — closer to where the experiment peaks — but it comes out clockwise, the wrong sign, with M̄ ≈ −0.10: a faint counterclockwise core wrapped in a clockwise outer-middle band. In open space a left veer does rotate counterclockwise (we check this directly: a lone walker in a large arena gives m > 0). Once confined, our inward wall response flips it — the net sign is set by how the veer meets the wall, not by the veer alone — so this minimal intrinsic model is not robust either.

In short: the experiment's coherent, disk-filling counterclockwise rotation does not fall out of either shortcut — a wall-only turn, or an independent constant veer under realistic heading noise. That is consistent with the paper's claim that the effect rests on a genuine individual locomotor bias whose collective expression a one-knob model does not casually recover. Reproducing the field, not just the mean, would need more than either minimal mechanism offers.

Running it

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# the full sweep (writes m_bar_sweep.csv, m_bar_table.csv, m_pdf.png):
python -m clockwise --fractions 0.0 0.25 0.45 0.7 1.0 --sizes 16 24 32 --seeds 10 --out study-output

# tests
pytest

Materials

The source paper and its data are in materials/ (see materials/README.md). This repository is a reproduction and test of that work, not original research; credit for the phenomenon and the experiments belongs to Echeverría-Huarte, Feliciani, Shi, Nishinari, Sánchez, Garcimartín & Zuriguel.