Learning the hierarchical structure of spatial environments using multiresolution statistical models

We explore the use of hierarchical Partially Observable Markov Decision Process (HPOMDP) models to represent and learn a multiresolution spatial structure representation of indoor office environments. The hierarchical POMDP model is based on the hierarchical Hidden Markov Model (HHMM). HPOMDPs can be learned from sequences of observations using an extension of the hierarchical Baum-Welch estimation algorithm for HHMMs. We apply the HPOMDP model to indoor robot navigation and show how this framework can be used to represent multiresolution spatial maps. In the HPOMDP framework the environment is modeled at different levels of resolutions where abstract states represent both spatial and temporal abstraction. We test our hierarchical POMDP approach using a large simulated (modeled after a real environment) navigation environment. The results show that the hierarchical POMDP model is more capable in inferring the spatial structure than a uniform resolution "flat" POMDP.