Robust Multi Agent Coordination: Integrating Causal Inference and Evolutionary Strategies in Adversarial Environments

Abstract

Autonomous multi-agent systems operating in safety-critical settings—bordersurveillance, infrastructure protection, maritime patrol, and contested reconnaissance—must sustain effective coordination even when actively targeted by adversaries. UAV swarms face a particular challenge: GPS spoofing, communication interference, and unexpected environmental disturbances can silently degrade individual agents before the rest of the swarm detects anything is wrong. This paper introduces a robust multi-agent coordination framework that tightly couples causal inference with evolutionary route selection inside a unified six-layer pipeline. A real-time threat detection and trust evaluation subsystem watches agent telemetry continuously, raises anomaly flags, and—critically—separates deliberate adversarial interference from ordinary environmental reactions using directed causal analysis. Agents identified as compromised are automatically quarantined; a hysteresis-based protocol then governs their gradual return to the swarm. Evolutionary path optimization adapts navigation for the remaining healthy agents, keeping the mission on track despite a degraded swarm. The framework runs in Unity 6 and was tested across 40 fully automated simulation runs covering seven distinct test categories. Detection reached 100%, every agent recovered successfully, and every mission completed, with a mean detection latency of 0.89 s, a mean recovery time of 14.56 s, and an evolutionary fitness score of 89.90. An ablation study shows the causal layer alone cuts false-positive isolation events by 85.7% compared to a simple threshold-only baseline.