Humans possess an exceptional ability to navigate and map their surroundings while continuously adapting to new environments. We effortlessly integrate prior knowledge with new sensory information, refining our spatial understanding over time. One of the long-standing challenges in robotics and Artificial Intelligence (AI) is developing a long-term Simultaneous Localization And Mapping (SLAM) system that can autonomously construct and update maps while adapting to dynamic and previously unseen environments. Despite significant progress in SLAM research, current state-of-the-art SLAM systems struggle with dynamic environments. Most SLAM algorithms are designed for static or predefined environments, and exhibit a lack of robustness when exposed to evolving or previously unseen conditions. Our best algorithms are still not robust enough to handle the complexity of a dynamic world, where factors such as lighting changes, moving objects, and environmental variations constantly challenge perception and mapping. These systems struggle to adapt to such changes over time, often leading to failures in localization and mapping consistency. Moreover, scalability for long-term navigation remains a challenge, as there are no definitive strategies for managing extremely large maps generated over hours or even days of navigation. Consequently, a true lifelong SLAM framework, capable of continuously adapting and updating its knowledge in large-scale environments has yet to be realized, leaving long-term autonomy in real-world scenarios an open challenge. In this dissertation, we focus on the integration of SLAM and continual learning to improve long-term localization and navigation in large-scale environments. Building on recent advances in machine learning and deep architectures for AI, we propose a framework that enables SLAM systems to continuously update their knowledge of the environment, adapting to dynamic and evolving scenarios, with a memory management mechanism that allows them to handle extensive navigations. Through this approach, we provide a foundation for more autonomous and reliable navigation in real-world scenarios.
Matteo Scucchia (Thu,) studied this question.