Open3D SLAM: Point Cloud Based Mapping and Localization for Education

Abstract

Modern LiDAR SLAM systems have shown remarkable performance and ability to operate in various environments ranging from indoor offices to large natural environments such as forests. This versatility has been afforded through many years of research that improved SLAM system components toward reliable real-time operation. However, achieving real-time computation has come at the cost of increased complexity and specific assumptions about the point cloud representation (e.g., LOAM and its variants). This extra complexity makes it more difficult for a non-expert or a student to dive into the field since extra effort is required to understand the ideas enabling real-time computation. On the other hand, the latter ideas often leave the underlying algorithmic principles unchanged. Furthermore, since SLAM performance is highly dependent on the implementation quality, the performance difference is often not caused by the underlying algorithm itself but rather by the implementation quality. Open3D SLAM tries to overcome these issues. We investigate using well-understood algorithms in their basic form to build the proposed LIDAR-based SLAM system. Our system leverages the Open3D library, which is well maintained and performant, thus contributing to the implementation quality and rendering Open3D SLAM open for future enhancements. Initial tests suggest that using basic algorithms as SLAM building blocks is viable on modern CPUs. We can build high-quality maps in different environments ranging from large outdoor scenes to small office environments. The generality of the proposed solution is demonstrated using different laser sensors deployed on various robotic platforms. We hope to make point cloud-based SLAM more accessible, thus facilitating teaching and enabling a new generation of mapping researchers to enter the field easier. Open3D SLAM will be used in the 3rd edition of ETH Robotic Summer School in July 2022, in Wangen a.d. Aare. The code is available on GitHub: https://github.com/leggedrobotics/open3d_slam