Researchers propose three-level learning architecture for autonomous UAV swarms in search and rescue

Researchers have proposed a novel three-level hierarchical learning architecture designed to enhance the autonomy of UAV swarms in search and rescue (SAR) operations. The architecture diverges from conventional approaches by integrating three distinct learning mechanisms aligned with a biological hierarchy: reflexes, skills, and reasoning.

At the reflex level, the system employs Hebbian neuroplasticity for individual agent adaptation, enabling rapid, localized responses to environmental stimuli. The skill level leverages multi-agent reinforcement learning (MARL) with graph neural networks (GNNs) and behavior trees to coordinate tactical maneuvers among swarm members. For strategic decision-making, the architecture incorporates model-agnostic meta learning with belief-desire-intention (BDI) reasoning and a digital twin to simulate and optimize long-term outcomes.

The design is formalized through 22 architectural contracts distributed across six components: BDI, Behavior Trees, GNN, MARL, Neuroplasticity, and Meta Learning. These contracts collectively provide six formal guarantees: safety, budget correctness, optimality, liveness, starvation freedom, and inter-level consistency. The authors introduce 'Swarm Meta Cognition' as a compositional property that emerges from the structured interaction of all three levels, allowing the swarm to monitor its own cognitive state and dynamically switch between cognitive strategies.

For dynamic scenarios involving active learning, the framework is extended with five additional contracts that introduce three new guarantees: cognitive resilience, graceful degradation, and monotonic meta improvement. The authors present a main integration theorem asserting that, when all contracts are satisfied, the hybrid neuro-symbolic system preserves all six guarantee classes. They further claim that the architecture addresses five fundamental limitations of existing hierarchical reinforcement learning approaches, though empirical validation is not provided in the paper.

The paper is published on arXiv under the cs.AI category and authored by Oleksii Bychkov. It includes a theoretical analysis but does not include experimental results or real-world deployments.