Grounded Communication Policies in Heterogeneous Agent Reinforcement Learning

Communication allows agents to overcome individual perceptual limitations, coordinate their behaviours, and operate collectively in multi-agent systems. Most existing approaches treat communication as an information-passing channel embedded within agents' action policies. As a result, communication is often abstracted as information exchange, whereas communication in humans typically involves establishing and maintaining common ground, that is, shared understanding of task-relevant information. This consideration becomes particularly relevant in heterogeneous-agent reinforcement learning (HARL), where agents differ in their perceptual or action capabilities. In this paper, I argue that communication in HARL should be modelled as a distinct, learnable process aimed at grounding shared beliefs, rather than as an auxiliary component of the action policy. I propose a framework in which each agent learns a separate communication policy that operates in parallel with its action policy to iteratively establish common ground before and during task execution. My work investigates how such grounded communication policies affect coordination, learning stability, and generalisation in HARL.