A Constructive Machine Learning Approach for Robot Exploration and Search

Semantic knowledge, more specifically semantic maps that associate semantic concepts (labels like ‘room’ and ‘corridor’) to spatial entities, has been employed to improve the performance of (multi-)robot planning tasks, such as search and exploration. However, although current semantic mapping approaches are very effective in labeling the parts of environments already visited by the robots, they are usually unable to predict the labels and, more generally, the structure of unvisited parts of environments. In this contribution, following a Constructive Machine Learning (CML) approach, we discuss the use of a generative method that is able to model and predict the topological structure and the labels of rooms for an indoor, previously unknown (or partially observed) environment. While this approach is not always able to find a perfect prediction of the structure of a given unknown environment, it seems nevertheless able to capture some fundamental structural properties. We explicitly note that the purpose of this paper is not to show any definitive results (although we provide a detailed example), but advocate the potential of using highlevel semantic knowledge to predict the structure of unknown parts of indoor buildings in order to improve exploration and search.