A Semi-supervised Method for Learning the Structure of Robot Environment Interactions

摘要

For a mobile robot to act autonomously, it must be able to construct a model of its interaction with the environment. Oates et al. developed an unsupervised learning method that produces clusters of robot experiences based on the dynamics of the interaction, rather than on static features. We present a semi-supervised extension of their technique that uses information about the controller and the task of the robot to (i) segment the stream of experience, (ii) optimise the final number of clusters and (iii) automatically select the individual sensors to feed to the clustering process. The technique is evaluated on a Pioneer 2 robot navigating obstacles and passing through doors in an office environment. We show that the technique is able to classify high dimensional robot time series several times the length previously handled with an accuracy of 91%.