Emergence of Grounded Language Representations for Continuous Object Properties Through Decentralized Embodied Learning

Yu Wei, Satoshi Nakagawa, Tomohiro Tanikawa, Yasuo Kuniyoshi

Abstract: While existing emergent language systems often lack robust grounding in real-world perceptual modalities, this study explores how environmental interventions and decentralized reinforcement learning frameworks can foster the emergence of grounded, adaptable communication protocols in multi-agent systems. By employing environment perturbations and data augmentation, agents develop the capacity to symbolize and exchange information about continuous attributes like color and spatial relationships. Analysis reveals that the emerged protocols capture grounded representations tied to the embodied environment, rather than merely exploiting statistical regularities. This research uncovers principles governing the emergence of communication protocols that flexibly represent and convey information in continuous perceptual domains. Unlike prior work focused on discrete concepts, our agents develop language grounded in embodied interaction, with evolved protocols exhibiting generalization that captures the underlying structure of continuous attributes. This work lays the foundation for developing intelligent agents with versatile linguistic capacities, enhancing human-agent collaboration in real-world settings with multi-dimensional continuous sensory data streams.

External IDs:doi:10.1007/978-981-96-0119-6_20