Resilient multi-agent reinforcement learning for centralised tactical conflict resolution under uncertain perturbations and non-cooperative traffic in urban air mobility

This research investigates tactical conflict resolution for Unmanned Aircraft Systems (UAS) and Urban Air Mobility (UAM) operations under degraded conditions and in the presence of non-cooperative UAS/UAM and manned Commercial Air Transportation and General Aviation (CAT/GA) intruders. The study adopts a centralised safety-net approach within UAS Traffic Management (UTM) architectures, envisioning ground-based conflict resolution services. We propose a set of Tactical Conflict Resolution Solvers (TCRS), each built upon a Multi-Agent Reinforcement Learning (MARL) core using a shared-policy transformer architecture and executed in a decentralised manner. To assess resilience of TCRS variants, we introduce domain-specific perturbations, including positioning noise, communication loss, and sensor-related defects. The TCRS operates with partial decision-making ability in non-cooperative traffic environments, while the perturbation model increases realism by simulating varying degrees of information availability. Results show that the perturbation-trained models achieve substantial safety gains compared with the baseline TCRS trained in ideal conditions. The most resilient variant; trained under multi-perturbation exposure and evaluated in non-cooperative environments, achieves a threefold reduction in critical safety violations compared with the baseline and remains robust under mixed cooperative/non-cooperative traffic with static intent. It exhibits a modest vulnerability under fully homogeneous non-cooperative scenarios with dynamic intent. Simulations involving concurrent CAT/GA and UAS operations further indicate that integrating UAS operations within the existing airspace classification remains hazardous for ground-based tactical conflict resolution when constrained by short look-ahead horizons and insufficient time to react.