Two Heads are Better than One: Collaboration-Oriented Multi-Agent Exploration System

Autonomous exploration in unknown environments is a complex and formidable challenge that requires effective collaboration among multiple agents under partially observable conditions. Due to limited observations and inefficient collaboration, multi-agent exploration often suffers from excessively long exploration paths. To address this issue, this paper proposes a Collaboration-Oriented Multi-Agent Exploration system (COMAE). To effectively understand and leverage the inter-agent relationships, this paper introduces Collaboration-Oriented Observation (COO). In addition to the basic connectivity graph, the COO further constructs collaboration-oriented node features and an interaction graph to enhance the overall strategic understanding of multi-agent. In order to improve collaboration among agents, this paper designs an Attention-based Sequential Network (ASN) to predict strategic actions. Additionally, a novel Collaborative Exploration Reward (CER) is proposed to further prevent non-collaborative behaviors during the exploration process. Extensive experiments demonstrate that the proposed method enhances collaboration among agents and significantly reduces exploration distances.