Listen, Adjust, Act: Adding Communication to Pre-Trained Agents via Goal Adjustments
Oriol Miro-Lopez-Feliu, Adrián Tormos, Víctor Giménez-Ábalos · Frontiers in artificial intelligence and applications · 2025
Effective coordination among intelligent agents is challenging, particularly in complex environments–often tackled with Multi-agent Deep Reinforcement Learning (MADRL). Communication is key to facilitate coordination, yet manually designing communication mechanisms is impractical. Instead, Comm-MADRL allows agents to learn meaningful communication without predefined semantics. However, conventional Comm-MADRL methods require jointly optimising communication and behaviour, which complicates training, often reducing its applicability to complex environments. Alternatively, we consider extending already trained agents without communication capabilities. In this paper we introduce a method that does so by extending pre-trained Goal-conditioned Reinforcement Learning (GCRL) agents, treating communication as modifications of the latent goal embeddings. The agent first trains in communication-less tasks, and then transfers its knowledge of the environment to tasks with communication. We evaluate the technique in a complex environment: MA-Minecraft, on two tasks that involve communication, showing significant performance improvements when coordination is required and inconclusive when it is helpful but optional. Our results suggest communication implemented as goal modifications in MADRL can bridge current methods towards richer, real-world-like MA scenarios.