MAGIC: Multi-Agent Generative Intention Coordination

David Huk, Oliver Hamelijnck, Dimitris Demiris, Theodoros Damoulas

Keywords: Multi-Agent Imitation Learning, Centralised Training Decentralised Execution, Coordination, Copula

TL;DR: We introduce MAGIC, a scalable framework for Multi-Agent Generative Intention Coordination that induces a structured joint policy without explicitly modelling the high-dimensional joint action space.

Abstract: Multi-agent imitation learning (MAIL) aims to learn coordinated policies from expert demonstrations, but learning \emph{dependent} multi-agent policies is often infeasible due to the combinatorial complexity of joint action modelling. As a result, many practical approaches assume a factorisation across agents, sacrificing expressivity and ignoring coordination present in expert data. We introduce MAGIC, a scalable framework for Multi-Agent Generative Intention Coordination that induces a structured joint policy without explicitly modelling the high-dimensional joint action space. \MAGIC follows a divide-and-conquer strategy: (i) we learn lightweight independent policies, (ii) we compress each agent’s action distribution into a one-dimensional latent \textit{intention} via either $\rho$- or Hilbert-space projections, and (iii) we learn a dependent generative model over intentions using diffusion-based copulas. This yields a scalable generative representation of the joint policy, enabling coordinated action sampling while preserving inter-agent dependencies. Moreover, in the centralised training with decentralised execution setting, \MAGIC supports coordinated execution without communication by allowing intention values to be precomputed offline. We show that \MAGIC outperforms MAIL baselines on a challenging real-world dataset.

Submission Number: 94