Abstract. Robot localization and mapping algorithms commonly represent the world as a static map. In reality, human environments consist of many movable objects like doors, chairs and tables. Recognizing that such environment often have a large number of instances of a small number of types of objects, we propose an alternative approach, Model-Instance Object Mapping that reasons about the models of objects distinctly from their different instances. Observations classified as short-term features by Episodic non-Markov Localization are clustered to detect object instances. For each object instance, an occupancy grid is constructed, and compared to every other object instance to build a directed similarity graph. Common object models are discovered as strongly connected components of the graph, and their models as well as distribution of instances saved as the final Model-Instance Object Map. By keeping track of the poses of observed instances of object models, Model-Instance Object Maps learn the most probable locations for commonly observed object models. We present results of Model-Instance Object Mapping over the course of a month in our indoor office environment, and highlight the common object models thus learnt in an unsupervised manner.