基于融合社交模型的多智能体强化学习导航方法OACHSSCD

Multi-agent cooperative navigation in dense crowds remains challenging due to complex interac-tion patterns and the difficulty of modeling social norms.To enhance collision avoidance efficiency and so-cial adaptability in such environments,this paper proposes SFMRVO-DRL,a multi-agent navigation framework that integrates Reciprocal Velocity Obstacles(RVO)with the Social Force Model(SFM)into a unified hybrid social interaction model.The framework adaptively balances the two components based on collision risk,enabling both efficient collision avoidance and naturalistic motion.Building on this hybrid interaction model,we design a decision architecture incorporating a graph attention mechanism to capture dynamic inter-agent relationships,and introduce a multi-objective reward function grounded in right-hand passing conventions to improve social compliance and policy stability.The proposed method employs MAPPO for policy optimization under a centralized-training and decentralized-execution paradigm.Experi-mental results in a canonical circular crowd-interaction scenario demonstrate that our approach significantly outperforms GA3C-CADRL,NH-ORCA,and HeR-DRL in terms of success rate,navigation time,and speed.This study highlights the effectiveness of hybrid RVO-SFM interaction modeling,attention-based social perception,and socially normative multi-objective reinforcement learning in advancing multi-agent navigation performance in complex crowd environments.