Multi-Robot Path Planning via an Adaptive Diffusion-Based Local Repair Method

This paper proposes a scalable multi-robot path-planning method designed for dense, obstacle-rich environments where conventional planners often incur substantial runtime overhead as the number of robots and constraints increases. The proposed approach integrates a diffusion-based generator with a classical planning framework and introduces budget-aware adaptive inference together with a targeted local repair mechanism. Instead of repeatedly invoking costly global replanning, the planner identifies critical regions along candidate solutions and performs local repairs to resolve conflicts and constraint violations efficiently, thereby improving responsiveness in high-density settings. Moreover, a computation-budget manager and experience-reuse strategy are employed to avoid redundant evaluations and to prioritize corrective updates where they yield the greatest benefit. Experiments across multiple challenging scenarios demonstrate that the proposed method maintains a high success rate and strong adherence to demonstration/data priors while significantly reducing planning time. Overall, the main contribution is a principled combination of adaptive inference scheduling and local repair that improves multi-robot planning efficiency without sacrificing solution quality, offering a practical and extensible basis for large-team robot coordination in complex environments.