Grid Map Fusion Method for Multiple Mobile Robots in Large-Scale Packaging and Printing Workshops

Abstract

To address the challenges in large-scale packaging and printing workshops, where the dense arrangement of electromechanical equipment leads to large-scale scenes and prominent repetitive structures, and where frequent dynamic disturbances impair the accuracy and efficiency of cooperative mapping by multiple mobile robots, this paper proposes a grid map fusion method that integrates multi-resolution cascading with optimal transport. The method constructs a coarse-to-fine pyramid hierarchy: at the coarse-resolution level, an initial estimate of the global alignment trend is obtained, and a cross-layer prior transfer mechanism maps the prior information to the finer level, with log-domain Sinkhorn iterations ensuring numerical stability in fine matching; at each resolution level, Oriented FAST and Rotated BRIEF (ORB) features are extracted, and a hybrid cost matrix is formulated by combining appearance information with local topological signatures, while virtual nodes are introduced to improve robustness against missing correspondences and dynamic outliers; high-confidence correspondence candidates are selected based on bidirectional normalized confidence and the maximum consensus criterion, and the Random Sample Consensus (RANSAC) algorithm is employed to estimate the two-dimensional rigid-body transformation; finally, a globally consistent grid map is generated through standard log-odds probability fusion. Experiments on a self-constructed packaging and printing workshop dataset demonstrate that the proposed method maintains registration accuracy while effectively reducing the computational burden of large-scale map fusion, thus providing a feasible technical pathway for global environment modeling and dynamic task scheduling in multi-robot cooperative operations.