Robot learning for loop closure detection and SLAM

Abstract

Robotics and autonomy continues to be a key research and development focus around the world. Robots are increasingly prevalent in everyday life. From manufacturing, home cleaning, to self-driving vehicles, robots are an ever-present reality with demonstrated ca- pability to increase quality of life for humans. As more and more robots exist surrounding humans, it becomes increasingly critical that robots can accurately sense and reason about the environment. The functionality of a robot building a map of its environment and lo- cating itself constantly within the map is known as Simultaneous Localization and Mapping (SLAM). SLAM is a difficult problem, and can be especially challenging when environmental appearance changers occur or when a GPS signal is not available. However, it’s within these challenging environments where the use of robots is critical. Consider a partially collapsed underground mine environment. If the environment is potentially dangerous, it doesn’t make sense to risk human life to enter the mine to perform search and rescue. If robots can be enabled to operate in challenging environments such as collapsed mines, human life can be saved. This Master’s thesis addresses the problem of increasing the effectiveness of SLAM in these challenging environments. First, I describe a data structure capable of capturing environmental metadata for semantic description overlay to augment mapping capability. Secondly, I introduce a novel loop closure detection technique that utilizes robot learning to understand complex environments. These efforts combined contribute to increasing the effectiveness of SLAM in GPS-denied environments or environments with varying lighting conditions.