Teaching agents to act socially.

Multi-Agent Path Finding (MAPF) asks a team of agents to find collision-free paths through a shared environment. Learning-based systems scale by deploying the same decentralized network to every agent—but identical policies can produce identical behavior, leaving agents locked in symmetric conflicts. SYLPH introduces dynamic social roles through Social Value Orientation (SVO). Each agent predicts its most influential interaction, selects an appropriate level of selfishness or altruism, and follows an SVO-conditioned motion policy. By sharing these social preferences with nearby agents, the team can produce nuanced, varied responses that break ties while retaining the scalability of parameter sharing. Across random, room-like, and maze-like maps, SYLPH outperforms prior learning-based MAPF planners and transfers to physical robots.