MR-FLOUR: Multi-robot Relative localization based on the Fusion of LiDAR, Odometry, and UWB Ranging

Abstract

Accurate relative localization of multiple robots is crucial for efficient collaboration and teaming, where a prior map of the environment is often unavailable. In this context, proximal robot detection plays an important role in improving relative localization accuracy by providing essential spatial awareness. While LiDAR is a common choice for detecting nearby robots, it struggles to distinguish them from surrounding obstacles, especially in cluttered environments. To address this challenge, we introduce MR-FLOUR, which stands for Multiple-robot Relative localization based on the Fusion of LiDAR detection outcomes, Odometry, and UWB Ranging. The main innovation of our approach is the use of different sensors for proximal robot detection and the introduction of our LiDAR detection constraint for optimization. First, we propose an efficient method to integrate UWB ranging with LiDAR data for proximal robot detection. We cluster the LiDAR point cloud and apply circle-fitting on the clusters based on the expected radius of the robot to reject clusters that do not conform to the expected shape of the robot. Then match the UWB ranging with cluster distances to determine nearby robot positions. Next, we estimate the identified robot’s orientation from successive detections, with outliers filtered using short-term odometry data. Finally, through Pose Graph Optimization (PGO), we fuse odometry and UWB ranging constraints with our proposed LiDAR detection constraint, which not only accounts for the position and orientation estimations of the nearby robots but also incorporates the relative pose estimation between them. Our method improves the localization accuracy of traditional UWB localization by incorporating LiDAR detection constraints when in Line-Of-Sight (LOS). In Non-Line-Of-Sight (NLOS) conditions or when no nearby robot detections are available, it relies on UWB and odometry for localization. We validated the approach with three robots in three indoor environments, achieving up to 33.3% improvement in translation and 45.5% in rotation over traditional UWB localization.