Localization for Indoor Robots using Implicit Map Representations

Accurate global localization is a fundamental capability for mobile robots in the indoor environment. It is the process of determining the robot's position and orientation within a known map, and is essential for many downstream tasks such as autonomous navigation, path planning, or manipulation. Indoor global localization presents a challenging task, especially in environments that frequently lack unique geometric or appearance cues, making it difficult to determine the robot's pose when multiple locations appear similar. Since external positioning systems are typically unavailable in indoor environments, robots must rely on onboard sensors such as light detection and ranging (LiDAR) sensors to localize themselves. To this end, the typical localization approach is to estimate the robot's pose by comparing sensor observations with a pre-built map representation of the environment. Therefore, the quality of the map representation is essential to localization performance. Traditionally, explicit geometric map representations such as occupancy grid maps have been the standard choice for indoor localization systems. However, these discrete representations suffer from the limitation that they require pre-defining a fixed resolution. This creates a trade-off between memory consumption and geometric accuracy, and introduces noise from the discrete grid structure. To address these limitations, implicit neural representations offer a promising alternative for encoding the environment as a continuous function parameterized by a neural network, which enables querying at arbitrary spatial locations and providing fine geometric details beyond resolution constraints. The main contributions of this thesis are novel approaches that exploit implicit neural representations to address the limitations of explicit geometric map representations and improve the performance of indoor robot localization using 2D LiDAR sensors. We first present a learning-based localization method that uses an implicit neural representation as an observation model for 2D LiDAR-based global localization. This approach exploits continuous implicit representations to enable accurate geometric modeling beyond the resolution constraints of discrete grid maps, providing superior performance in challenging scenarios with repetitive structures. Based on this approach, we achieve state-of-the-art localization performance compared to approaches using explicit map representations and successfully localize robots in all test scenarios, while several baseline methods fail to converge in challenging scenarios. While this approach demonstrates the advantages of implicit representations over explicit map representations, it faces computational challenges when dealing with large-scale environments and a large number of particles. To address this limitation, we further propose an efficient implicit map representation that significantly improves computational efficiency while maintaining superior localization accuracy, making it highly suitable for real-world applications.