A Mechanistic Model for Collective Motion from Sensorimotor Regularities

Technical Summary: A Mechanistic Model for Collective Motion from Sensorimotor Regularities

Problem Statement
Collective behavior in animals (e.g., flocks, schools, swarms) has traditionally been modeled using self-propelled particle (SPP) frameworks. While SPP models successfully reproduce group-level phenomena like polarization and milling through abstract interaction forces, they are fundamentally descriptive rather than mechanistic. Their parameters (interaction radii, alignment strengths) are fitted to observed patterns and lack correspondence to measurable biological quantities. Furthermore, SPP models often assume agents have direct access to neighbors' positions and headings, an assumption contradicted by recent empirical work (e.g., Sayin et al.) showing that sensory and cognitive mechanisms, rather than direct alignment, mediate interaction. There is a need for a model grounded in the sensorimotor level—operating on what organisms can actually perceive, estimate, and physically execute—to explain how collective behavior is produced and how it varies across species.

Methodology
The authors present a mechanistic model based on the Active InterCONnect (AICON) framework, extended from robotics to collective motion. Unlike rule-based or policy-driven approaches, AICON generates behavior through the dynamic composition of sensorimotor regularities.

• Architecture: Each agent maintains recursive probabilistic estimators for its own pose and, for each neighbor, for position and visibility.
• Perception: Agents perceive neighbors via two angular cues within a limited field of view ($\psi$): bearing ($\phi$) and apparent size ($\gamma$). These encode directional and distance information without requiring explicit depth sensing.
• State Estimation: Agents maintain uncertain internal state estimates. When a neighbor leaves the field of view, the estimator propagates the last belief forward, accumulating uncertainty (memory). If memory is disabled, estimates are discarded upon re-entry.
• Action Selection: Agents do not follow prescribed interaction forces. Instead, they minimize a single differentiable cost function: the expected deviation from a desired social distance ($d_0$) across all estimated neighbors.
  $$g_i = \sum_{j \in V_i} E[| \|x_j\| - d_0 |]$$
  Through gradient descent on this cost, three distinct behavioral gradients emerge naturally:
  1. Position Adjustment: Moving to reduce distance error.
  2. Active Sensing: Moving to reduce estimation uncertainty (triangulation).
  3. Reorientation: Turning to maintain neighbors within the field of view to preserve visibility and reduce uncertainty.
• Constraints: Agents are subject to physical turning constraints ($|\omega| \le \omega_{max}$), representing turning agility.

Experimental Setup
Simulations were conducted with $N=250$ agents in a 2D continuous environment using unicycle kinematics. The study varied five key sensorimotor parameters (Table 1):

1. Field of View ($\psi$): $\{90^\circ, 180^\circ, 270^\circ, 360^\circ\}$
2. Turning Agility ($\omega_{max}$): $\{0.1, 0.3, 0.6, 0.9\}$
3. Bearing Noise ($\sigma_\phi$): $\{0.001, 0.01, 1, 10, 100\}$
4. Apparent Size Noise ($\sigma_\gamma$): $\{0.001, 0.01, 1, 10, 100\}$
5. Process Noise ($Q$): $\{0.01, 0.1, 1, 10\}$

Behavior was quantified using metrics for polarization, spatial compactness (Relative Circular Area), global displacement, cohesion, and directional fragmentation. A global variance-based sensitivity analysis (Sobol indices) was performed to determine parameter influence.

Key Results
The model demonstrates that a diverse range of collective behaviors emerges solely from variations in sensorimotor parameters, without changing the underlying interaction architecture:

• Emergent Behaviors:
  • Stable Rings: Occur at low uncertainty with high turning agility, where agents maintain reliable estimates and accurate distance control.
  • Milling: Emerges under increased uncertainty. Moderate uncertainty produces loose irregular rings; high uncertainty collapses agents into dense, rotating clusters.
  • Fragmentation: Narrow fields of view combined with high noise and no memory cause fragmentation into drifting groups or parallel streams, as agents cannot maintain estimates of occluded neighbors.
• Parameter Influence:
  • Turning Agility ($\omega_{max}$): Governs formation geometry (ring radius) when estimation is reliable but only affects cluster tightness under high uncertainty.
  • Sensory Noise: Bearing noise disrupts directional localization (disordered aggregation), while apparent size noise degrades distance estimation (inward-spiraling formations). These produce distinct behavioral signatures, unlike scalar noise models.
  • Memory: Critical for cohesion under narrow fields of view. Its absence leads to qualitative fragmentation even with identical initial conditions.
• Sensitivity Analysis: Field of view ($\psi$) and process noise ($Q$) are the dominant direct drivers of collective behavior. Sensory noise parameters act primarily through interactions (high total-order indices). Memory redistributes sensitivity, making the system less dependent on any single parameter.

Significance and Claims
The paper claims to provide a constructive alternative to descriptive particle models by grounding collective behavior in sensorimotor regularities.

• Mechanistic vs. Descriptive: The model explains how behavior is produced rather than just what it looks like. Parameters correspond to measurable biological quantities (e.g., turning agility, sensory noise, field of view geometry) rather than fitted constants.
• Embodiment and Environment: Differences in collective behavior across species are framed as emergent outcomes of differences in embodiment (sensorimotor properties) and environment, rather than differences in abstract interaction rules.
• Active Sensing: The model naturally produces active sensing behaviors (triangulation, reorientation) as a consequence of minimizing uncertainty within the sensorimotor loop, a feature absent in memoryless response models or abstract Bayesian models lacking geometric grounding.
• Predictive Power: The framework suggests that behavioral transitions and coordination breakdowns can be predicted by manipulating measurable biological parameters, offering a path to transfer predictions to new species without refitting.

Limitations
The authors note that cross-species predictions remain a hypothesis requiring validation against empirical data. The current model assumes all agents are identical and treats neighbor interactions as globally contributing to the cost function regardless of occlusion (though visibility is modulated), which may constrain observable structures. Future work is suggested to include geometrically localized interactions, individual variation, and structured environmental perturbations.