Abstract
Early cavemen relied on gestures, vocalizations, and simple signals tocoordinate, plan, avoid predators, and share resources. Today, humanscollaborate using complex languages to achieve remarkable results. What drivesthis evolution in communication? How does language emerge, adapt, and becomevital for teamwork? Understanding the origins of language remains a challenge.A leading hypothesis in linguistics and anthropology posits that languageevolved to meet the ecological and social demands of early human cooperation.Language did not arise in isolation, but through shared survival goals.Inspired by this view, we investigate the emergence of language in multi-agentForaging Games. These environments are designed to reflect the cognitive andecological constraints believed to have influenced the evolution ofcommunication. Agents operate in a shared grid world with only partialknowledge about other agents and the environment, and must coordinate tocomplete games like picking up high-value targets or executing temporallyordered actions. Using end-to-end deep reinforcement learning, agents learnboth actions and communication strategies from scratch. We find that agentsdevelop communication protocols with hallmark features of natural language:arbitrariness, interchangeability, displacement, cultural transmission, andcompositionality. We quantify each property and analyze how different factors,such as population size and temporal dependencies, shape specific aspects ofthe emergent language. Our framework serves as a platform for studying howlanguage can evolve from partial observability, temporal reasoning, andcooperative goals in embodied multi-agent settings. We will release all data,code, and models publicly.