Social Touch as an Allostatic Tool in Cooperative Multi-Agent Reinforcement Learning

This study explores the computational modeling of social touch in a cooperative multi-agent reinforcement learning (MARL) setting.Inspired by research on social allostasis and affective neurobiology, we implement a novel model of social touch in artificial agents within a cooperative game setting, where interoception-inspired internal states are modulated by eating and agent-to-agent contact.A memory-augmented Proximal Policy Optimization (PPO) is used to explore and exploit social touch for cooperative reward optimization and internal state regulation.Our paradigm reveals that agents trained with modulatory social touch achieve better game performance, even in environments where touch is ineffective.We observe an evolution of touch behavior from frequent exploration to selective usage, showing better internal-state regulation and adaptability to heterogeneous starting conditions between agents.The findings demonstrate how interactive systems might incorporate non-verbal communication cues to enhance cooperation, with implications for designing more intuitive human-AI and robot-robot collaborative systems.