SIGMAS: Second-Order Interaction-based Grouping for Overlapping Multi-Agent Swarms
New self-supervised framework decodes overlapping drone fleet behavior by analyzing second-order interactions.
Researchers Minah Lee and Saibal Mukhopadhyay have introduced SIGMAS (Second-order Interaction-based Grouping for Multi-Agent Swarms), a novel AI framework published in a paper accepted at AAMAS 2026. The system addresses the critical challenge of understanding complex swarm systems—such as coordinated drone fleets or robotic teams—where traditional analysis methods struggle. Unlike pedestrian crowds or traffic, swarms feature persistent group memberships that are essential to decode for predicting behavior. SIGMAS pioneers the task of group prediction in overlapping swarms, requiring it to infer these latent structures purely from observed movement trajectories, without any supervised training data telling it what groups exist.
The technical breakthrough lies in SIGMAS moving beyond simple pairwise interactions to model second-order relationships: it analyzes not just if two agents interact, but how similarly each agent interacts with all others in the system. This allows the framework to capture higher-order social structures. A key component is a learnable gating mechanism that dynamically balances the influence of individual agent behavior versus emergent collective dynamics during reasoning. In diverse synthetic swarm experiments, SIGMAS demonstrated robust performance in accurately recovering the true, hidden group assignments, even when swarm dynamics overlapped simultaneously. This work establishes both a new benchmark task and a principled, self-supervised modeling approach that could significantly improve autonomy and coordination in real-world robotic systems.
- SIGMAS is a self-supervised AI framework that infers latent groups in swarms (e.g., drone fleets) from raw trajectory data alone, requiring no labeled training data.
- It models second-order interactions—analyzing how similarly agents interact with others—going beyond direct pairwise relationships to understand complex swarm dynamics.
- The system includes a learnable gating mechanism for adaptive reasoning and was validated across synthetic scenarios, accurately recovering group structures even under overlapping conditions.
Why It Matters
Enables better prediction and control of autonomous drone fleets and robotic teams by understanding their inherent social structures from movement alone.