Bayesian-Enhanced hierarchical reinforcement learning for robust multi-sweeper task scheduling
Yiwen Huang, Lan Juncong, Qinzhang Sun, Qian Cong, Shaofeng Lu · Expert Systems with Applications · 2026
To address the challenges of high scheduling complexity and insufficient robustness resulting from the uncertainty of task information in multi-unmanned sweeper systems within large-scale urban road networks, this paper proposes a hierarchical scheduling method based on Bayesian Multi-Agent Reinforcement Learning. The method utilizes a hierarchical architecture to decouple the global scheduling problem into upper-level task allocation and lower-level path planning, and improves computational efficiency and decision-making quality in complex scenarios through an interaction mechanism between the upper and lower layers. The core innovation lieslein the MAPPO-Bayesian algorithm designed for the upper layer, which analytically updates the posterior distribution of the value function via Bayesian linear regression. By quantifying the agents’ epistemic uncertainty, it guides exploration and decision-making. Concurrently, the lower layer generates optimal cleaning paths by incorporating vehicle constraints and task requirements, thereby achieving collaborative optimization. Experimental results demonstrate that, compared with MAPPO, QMIX and VDN, the proposed method reduces the coefficient of variation to 0.28, significantly enhancing training stability. Simultaneously, this architecture ensures that the multi-vehicle system maintains at least 95% effective working time during task execution, realizing efficient collaboration. Finally, the effectiveness of the proposed method is validated on the CommonRoad simulation platform developed by the Technical University of Munich, utilizing a structured road network of the Guangzhou International Campus of South China University of Technology constructed via OpenStreetMap. This study not only addresses the task scheduling problem in the field of unmanned cleaning but also provides a novel algorithmic framework for multi-agent collaborative systems.