Enhancing heterogeneous multi-agent autonomous exploration in unstructured terrains via dual-layer collaborative...

Multi-Agent Systems (MAS) are of great significance in real-world applications, as they can effectively manage complex tasks through decentralized collaboration. Most existing studies predominantly emphasize homogeneous MAS, i.e., systems consisting of agents with identical functional capabilities and standardized behavioral protocols. However, such systems often fail to adapt effectively to dynamic or disturbed environments during autonomous cooperative exploration. To this end, this paper investigates heterogeneous MAS and introduces a Reinforcement Learning-guided heterogeneous Multi-agent Autonomous Exploration framework (RL-MAE), designed to harness the complementary advantages of diverse agent types. Specifically, the LEP module achieves dynamic distortion pose compensation through a motion correction strategy tailored for heterogeneous platforms, providing a reliable perceptual foundation for task allocation and control. Furthermore, the HDA module establishes a three-level closed-loop architecture spanning planning, navigation, and adaptive coordination, where the upper layer determines target allocation and global paths via MD-MTSP, the middle layer performs local path tracking using TD3, and the lower layer triggers trajectory adjustments via dual-mode cost evaluation only when a cost-reduction condition is satisfied, with the three levels tightly coupled through explicit information feedback. Extensive experiments, including ablation studies, demonstrate that the proposed system achieves superior robustness compared to homogeneous systems in complex unstructured terrain.