Goal-oriented multi-robot collaborative source search with dynamic exploration–exploitation balance in large-scale constrained areas

Rapid localization of unknown gas leak sources in urban areas is critical for effective emergency response and impact mitigation. While deploying autonomous robots to assess and localize emission sources has proven effective, current approaches are inadequate in a large-scale, constrained area. To address this, we propose a Goal-oriented multi-Robot collAborative Source Search (GRASS) framework for large-scale constrained environments. This framework employs a three-step coupled strategy—goal determination, allocation, and execution—leveraging the fusion of posterior probabilities and hybrid sensed information to achieve efficient and reliable gas source localization. Specifically, goal determination module defines two distinct goals to dynamically balance exploration (reducing estimation uncertainty through systematic coverage) and exploitation (directing robots toward estimated source locations via Gaussian Mixture Models). Moreover, goal allocation module adapts to source estimation reliability and local context, enabling robots to prioritize exploration during sparse data periods and shift to exploitation as estimations improve. In terms of goal execution, it resolves conflicts between goal pursuit and collision avoidance through designing adaptive path planning mechanisms that integrates a modified A* algorithm with a contour-tracing method. Finally, extensive simulations demonstrate that GRASS significantly outperforms two baseline methods, achieving higher success rates (increasing at least 9%) and requiring less search time (reducing at least 215.95 s) in various settings. These advantages are also confirmed by a real-world case study. Our work advances information fusion-driven environmental monitoring for resilient cities by providing an autonomous solution for source localization in complex urban environments.