Multi-Agent Path Finding (MAPF) is a core technology for logistics automation in factories and warehouses. Guidance-based approaches that reflect the structural properties of factory environments have been widely adopted for computational efficiency and execution feasibility. However, these approaches generally assume static environments in which predefined guidance policies remain valid. Therefore, unexpected obstacles can cause inter-robot collisions or deadlocks. To make nominal MAPF plans robust against execution uncertainty, prior studies have incorporated bounded execution delays into MAPF. A representative method is k Robust Multi-Agent Path Finding (kR-MAPF), which models allowable execution delay using a global robustness parameter k. However, when large obstacle-induced delays are represented by a single global robustness parameter, kR-MAPF imposes unnecessary conservatism and increases the search space. This increase in search space raises planning runtime and reduces path efficiency in large-scale robot fleet operation. This paper proposes a multi-robot path planning framework that updates guidance policies for each segment based on real-time obstacle information. The proposed framework identifies robots affected by obstacles and selectively replans their paths, thereby reducing unnecessary computation while maintaining path planning success and path efficiency. Simulation results in a 100m×100m factory environment with up to 100 robots demonstrate that the proposed framework maintains a 100% success rate under all tested conditions. Compared with kR-MAPF with different values of k, the proposed framework reduces planning runtime by approximately 35–79% and flowtime by approximately 7–24%. These results demonstrate that obstacle-aware selective replanning can improve both real-time performance and path efficiency in dynamic factory environments. The proposed framework provides a technical basis for stable large-scale multi-robot operation in structured industrial environments.
Park et al. (Wed,) studied this question.