Efficient resource scheduling for modular hospital fit-out accelerates the deployment of medical facilities, ultimately determining the project inauguration timeline. However, fit-out scheduling faces the real-time coordination of multiple rooms and workers, complex matching relationships between multi-skilled workers and diverse operations, and the dynamic accumulation of worker fatigue. These factors lead to varying evolutions of worker processing times, worker-operation affinities, and fatigue accumulation-recovery patterns across execution environments, causing temporal state-trajectory out-of-distribution (OOD) shifts between training and test settings. Although deep reinforcement learning holds potential for dynamic scheduling, existing methods struggle to adequately model dynamic relationships among resources and heavily rely on Markov assumptions, neglecting the cumulative fatigue effects in historical states. Therefore, this paper proposes a Heterogeneous Resource Affinity Modeling and Memory-Enhanced Deep Reinforcement Learning framework (HAMMER). It first introduces a resource affinity state modeling network based on a two-stage heterogeneous graph structure to dynamically characterize the compatibility between workers and operations. Furthermore, a temporal convolutional gated attention network is proposed to capture the long-term cumulative fatigue effects. Built on the Actor-Critic architecture and optimized with proximal policy optimization, HAMMER synthesizes these features to generate adaptive decisions. Experimental results show that HAMMER consistently improves fit-out efficiency, achieving an 8.3% average makespan reduction over the strongest DRL baseline. Shifted-dataset tests confirm its OOD generalization under altered processing-time distributions and fatigue parameters, while ablation studies validate the contributions of graph-based affinity modeling and memory-enhanced fatigue representation. Note to Practitioners—This study is motivated by the resource scheduling challenges in the modular hospital fit-out process. In this process, numerous medical function rooms and tight timelines require workers to operate across multiple rooms simultaneously. Efficiently allocating multi-skilled workers while balancing workload and fatigue is crucial to ensuring timely delivery and maintaining fit-out quality. Traditional scheduling methods often rely on experience or static rules, making it difficult to adapt to the real-time changes in fit-out progress and worker fatigue conditions. To address these challenges, this paper proposes the HAMMER framework. This intelligent scheduling solution identifies the operational suitability between workers’ skill levels and fit-out tasks, while capturing the long-term cumulative effects of worker fatigue on processing productivity. This enables adaptive and proactive resource scheduling under dynamic fit-out environments. Experiments on multiple simulated fit-out projects of varying scales demonstrate that HAMMER can significantly reduce project makespan and effectively control worker fatigue, showing higher robustness and efficiency compared with rule-based and traditional optimization methods. From an engineering perspective, HAMMER can be integrated into existing digital construction management systems or BIM platforms, where it can utilize sensor data and progress monitoring to enable dynamic task planning and resource optimization. Beyond modular hospital fit-out scheduling, the framework can also be extended to other labor-intensive industrial scenarios, such as manufacturing workshop scheduling. However, the current method requires sufficient real-time data and involves high computational costs in large-scale scheduling applications. In future work, we aim to develop a lightweight, distributed scheduling framework to enhance HAMMER’s efficiency and scalability across projects of different sizes.
