| dc.description.abstract | In this thesis, we propose novel approaches to robust multi-robot Simultaneous Localization and Mapping (SLAM). We explore multi-robot SLAM in both the centralized and distributed setting, with point-cloud or mesh-based representations to create a lightweight but dense map of the environment from vision or lidar data.
We present and discuss four different approaches to multi-robot SLAM. The first approach, named Large-scale Autonomous Mapping and Positioning (LAMP), is a centralized lidar-based approach that creates a dense point-cloud map of the environment and uses incremental Pairwise Consistency Maximization (PCM) to reject outliers in the loop closure measurements. The second approach, named meshLAMP, is an extension of LAMP that uses Graduated Non-Convexity to reject ouliers and creates a lightweight mesh map of the environment. To address the limitations in communication range and scalability of the centralized approaches, we also present two distributed approaches. The first distributed approach, named Distributed, Online, and Outlier Resilient SLAM (DOOR-SLAM), is a vision-based distributed approach to multi-robot SLAM that extends PCM to reject outliers without relying on centralized computation. The last approach, named Kimera-Multi, is a visionbased distributed approach that uses PCM for outlier rejection, and extends the lightweight mesh-based mapping module in meshLAMP to operate in a distributed fashion and generate a semantic mesh map of the environment. We demonstrate the four approaches in a variety of conditions, from indoor environments to photo-realistic simulators, to underground spaces in the context of the DARPA Subterranean Challenge, and show that they are able to perform reliably in the field. We conclude by commenting on the advantages and possible improvements for each approach. | |
