Compositionality in multi-robot AI

Rather than training a single large model to handle every task, we study how to decompose robot intelligence into specialized, reusable modules that can be composed at runtime. This compositionality-first perspective enables systems that generalize to novel task combinations, scale more efficiently than monolithic alternatives, and are easier to audit and retrain. Our work connects modular neural architectures with multi-agent coordination theory, grounding compositional principles in real hardware experiments with drone swarms and heterogeneous robot teams.