Skip to Main Content
Ambiguity in sensory-motor signals from a mobile robot due mainly to noise and fluctuation makes a deterministic approach unsatisfactory. In this paper, a stochastic approach based-on Hidden Markov Models (HMMs) is proposed to recognize environment of a mobile robot. From this recognition a graph-based map is formed. Graph-based maps are important in decreasing memory and the computational cost. Two methods for constructing graph-based maps are proposed. The former is to estimate HMMs based on quantized sensory-motor signals. The latter is to estimate HMMs based on a sequence of labels obtained by modular network SOM (mnSOM). Although mnSOM learns non-linear dynamics of sensory-motor signals, it still generates labels from each subsequence separately. This might not be robust, because resulting sequence of labels may rapidly change, which rarely occurs in the real world. This motivates us to combine mnSOM and HMM to realize more robust segmentation of the environment. The resulting HMMs can be converted into a graph-based map in a straightforward way. The resulting graph-based map is also useful for goal seeking. Simulation results demonstrate that the proposed method can construct graph-based maps effectively, and can perform goal seeking in the changing environment.