Modular Mapping and Learning for Robotics

Abstract

The ability to precisely localize a robot within the environment is a core capability for mobile robotics. Knowing a robot’s precise pose permits various tasks, such as navigation in an environment, interaction between multiple agents, or mobile manipulation of an object. Various applications such as, for example, autonomous driving, delivery robots, augmented reality (AR), mobile manipulation, and robot inspection greatly benefit from an accurate and robust underlying Simultaneous Localization And Mapping (SLAM) system. Motivated by these widespread applications, the goal of this thesis is to improve the existing limits of SLAM. We focus on integrating deep learning into various components of SLAM, as we aim to integrate the advances in perceptual understanding deep learning has provided. We wish to draw inspiration from how humans perceive the environment to include more semantically meaningful landmarks in the mapping process. We believe deep learning will not only improve existing solutions but provide novel solutions to yet unsolved problems. Different robots and environments have different optimal sensor configurations and perception requirements regarding SLAM. For example, micro aerial vehicles have limited payload and a different perspective than indoor wheeled robotic platforms. Accommodating these differences in software requires a flexible framework that can support multiple different modalities and types of sensors. However, many existing frameworks are very rigid and require significant changes when prototyping new products or testing novel research ideas. Therefore, in the first part of the thesis, we will focus on a new multi-modal, multi-robot, and modular SLAM framework. We show through experiments that state-of-the-art accuracy is possible, and we showcase the integration of various experimental modules. In the second part of the thesis, we will focus on using deep learning to improve SLAM, first, by working on a new type of landmark and descriptor. We further explore the inclusion of semantics into the localization process of our proposed framework. Finally, we use learning to address maintaining accurate long-term calibration in SLAM, which is essential to the operation of all downstream tasks. We focus on the modular and meaningful integration of learning, as we believe this leads to robust robotic systems that remain understandable and easier to design.