Multi-Agent Embodied Visual Semantic Navigation with Scene Prior Knowledge