Long-term robot mapping in dynamic environments

Abstract

One of the central goals in mobile robotics is to develop a mobile robot that can construct a map of an initially unknown dynamic environment. This is often referred to as the Simultaneous Localization and Mapping (SLAM) problem. A number of approaches to the SLAM problem have been successfully developed and applied, particularly to a mobile robot constructing a map of a 2D static indoor environment. While these methods work well for static environments, they are not robust to dynamic environments which are complex and composed of numerous objects that move at wide-varying time-scales, such as people or office furniture. The problem of maintaining a map of a dynamic environment is important for both real-world applications and for the advancement of robotics. A mobile robot executing extended missions, such as autonomously collecting data underwater for months or years, must be able to reliably know where it is, update its map as the environment changes, and recover from mistakes. From a fundamental perspective, this work is important in order to understand and determine the problems that occur with existing mapping techniques for persistent long-term operation. The primary contribution of the thesis is Dynamic Pose Graph SLAM (DPG-SLAM), a novel algorithm that addresses two core challenges of the long-term mapping problem. The first challenge is to ensure that the robot is able to remain localized in a changing environment over great lengths of time. The second challenge is to be able to maintain an up-to-date map over time in a computationally efficient manner. DPG-SLAM directly addresses both of these issues to enable long-term mobile robot navigation and map maintenance in changing environments. Using Kaess and Dellaert's incremental Smoothing and Mapping (iSAM) as the underlying SLAM state estimation engine, the dynamic pose graph evolves over time as the robot explores new areas and revisits previously mapped areas. The algorithm is demonstrated on two real-world dynamic indoor laser data sets, demonstrating the ability to maintain an efficient, up-to-date map despite long-term environmental changes. Future research issues, such as the integration of adaptive exploration with dynamic map maintenance, are identified.