Mobile robots have become a mature technology. The first cable guided logistics robots were introduced in the industry almost 60 years ago. In this time the market for mobile robots in industry has only experienced a very modest growth and only 2.100 systems were sold worldwide in 2011. In recent years, many other domains have adopted the mobile robots, such as logistics robots at hospitals and the vacuum robots in our homes. However, considering the achievements in research the last 15 years within perception and operation in natural environments together with the reductions of costs in modern sensor systems, the growth potential for mobile robot applications are enormous. Many new technological components are available to move the limits of commercial mobile robot applications, but a key hindrance is reliability. Natural environments are complex and dynamic, and thus the risk of robots misinterpreting the environment or failing to detect critical circumstances is unavoidable. To deal with this challenge, the control of robot applications must be able to handle imperfect observations and gracefully recover from unavoidable errors. The controller needs to know what is going on. This thesis addresses exactly this problem from the hypothesis that an assessment of the situation for the robot, will be able to contribute with essential knowledge to the robot control and enable the understanding of the current situation as well as predict the future status. A novel framework for situation modeling are presented, which applies an Extensible Markov Model (EMM) to represent the spatio-temporal nature of situations. On-line data-streams from the robot sensors and algorithms are processed using stream-based clustering to build the spatio-temporal structure or match the situation of the robot to existing states. Situation prediction is proposed using an on-line graph-search of maximum likelihoods in the EMM. The developed software modules are integrated in a new software architecture, which facilitates integration into any robotic control framework and uses on-line visualization of the spatio-temporal graphs to optimize situation classification. The results are evaluated in three real-world scenarios, which successfully evaluates capabilities of the proposed situation assessment framework within detection of known spatio-temporal relations, deviation from known spatio-temporal patterns, and detection of known critical situations.