Utilizing service robots in real-world tasks can significantly improve efficiency, productivity, and safety in various fields such as healthcare, hospitality, and transportation. However, integrating these robots into complex, human-populated environments for continuous use is a significant challenge. A key potential for addressing this challenge lies in long-term modeling capabilities to navigate, understand, and proactively exploit these environments for increased safety and better task performance. For example, robots may use this long-term knowledge of human activity to avoid crowded spaces when navigating or improve their human-centric services.This thesis proposes comprehensive approaches to improve the mapping, localization, and task fulfillment capabilities of service robots by leveraging multi-modal sensor information and (long- term) environment modeling. Learned environmental dynamics are actively exploited to improve the task performance of service robots.As a first contribution, a new long-term-autonomous service robot is presented, designed for both inside and outside buildings. The multi-modal sensor information provided by the robot forms the basis for subsequent methods to model human-centric environments and human activity.It is shown that utilizing multi-modal data for localization and mapping improves long-term robustness and map quality. This especially applies to environments of varying types, i.e., mixed indoor and outdoor or small-scale and large-scale areas.Another essential contribution is a regression model for spatio-temporal prediction of human activity. The model is based on long-term observations of humans by a mobile robot. It is demonstrated that the proposed model can effectively represent the distribution of detected people resulting from moving robots and enables proactive navigation planning.Such model predictions are then used to adapt the robot’s behavior by synthesizing a modular task control model. A reactive executive system based on behavior trees is introduced, which actively triggers recovery behaviors in the event of faults to improve the long-term autonomy. By explicitly addressing failures of robot software components and more advanced problems, it is shown that errors can be solved and potential human helpers can be found efficiently.