Topological approaches for simultaneous localization and mapping

摘要

Recently, various types of robots have developed for the exploration and monitoring in unknown and dynamic environment. Especially, the expectation of robots used in disasters is increasing to prevent the second disaster. Especially, it is very important to extract the environmental information related for remote control and monitoring of mobile robots. Simultaneous localization and mapping (SLAM) is an important methodology to deal with environmental information. Various types of methods for SLAM have been proposed such as Extended Kalman Filter (EKF) SLAM, Graph SLAM, visual SLAM, and cooperative SLAM. In general, there are two main approaches of grid mapping and topological mapping in the study on SLAM. In this talk, we focus on topological mapping methods to extract environmental features from a 3D point cloud. Various types of unsupervised learning methods based on topological mapping have been proposed to deal with environmental features in unknown environments. One of them is Growing Neural Gas (GNG) that can dynamically change the topological structure composed of nodes and edges. The advantage of GNG is in the incremental learning capability of nodes and edges according to target data distribution. We have proposed several different topological mapping methods based on GNG to extract the environmental features from a 3D point cloud until now. In this talk, we explain the research background of SLAM, the learning algorithm of GNG, and experimental results of GNG for SLAM in various environments. Next, we explain multi-layer GNG to extract hierarchical features in environmental maps as a multi-scale approach, and batch learning algorithm for GNG (GL-GNG) to improve the convergence property. Furthermore, we explain the modified method of GNG-utility (GNG-U), that we called GNG-U2. GNG-U2 can improve the real-time adaptability of extracting topological structure in non-stationary data distribution. Next, show experimental results of SLAM based on GNG-U2 in dynamic environments. Finally, we show several other application examples of topological approaches, and discuss the applicability and future direction of topological approaches in robotics.