LAUSR

Collaborative perception in unknown environments is a critical task for multi-robot systems. Without external positioning, multi-robot mapping systems have relied on the transfer place recognition (PR) descriptors and sensor data for the relative pose estimation (RelPose) and share their local maps for collaborative mapping. Thus, in a communication limited environment, data transmission can become a significant communication bottleneck in the multi-robot mapping system. Although a Gaussian Mixture Model (GMM) map and a submap-based framework have been proposed to reduce map data transmissions, the PR descriptors and sensor data for RelPose consume much of the communication bandwidth. Furthermore, the previous GMM submap construction methods may fail the multi-agent RelPose due to inconsistent weights. With a fixed number of Gaussian components, GMM submaps also have a limited ability to adapt to drastic changes in environmental characteristics while exploring. To address these limitations, this paper presents MR-GMMapping, a Multi-Robot GMM-based mapping system in which robots only transfer GMM submaps. We propose a novel GMM submap construction strategy with an adaptive model selection method, which can dynamically select the appropriate Gaussian model during exploration. Experiments on both simulators and real robots show that MR-GMMapping improves the accuracy of RelPose by 11% in average translation error and 30% in average rotation error in comparison with not GMM-submap-based method. In addition, data transmissions between robots are reduced by 98% in comparison to point cloud maps. MR-GMMapping is published as an open-source ROS project at https://github.com/efc-robot/gmm_map_python.git.