Application of Unsupervised Clustering to Complex Robot Training Tasks

The interest in obtaining sensor-motor competences by robot training – by expressing the desired actions of the robot in terms of its sensory perception using non-linear mapping techniques – has been growing steadily in mobile robotics applications. However if the sensor-motor task under investigation is subject to state transitions – in the sense that the casual relationship between sensor perception of the robot and the desired motor commands exhibits different characteristics over time or space -, it may not be possible to identify the whole sensor-motor relationship in a single model using standard non-linear mapping techniques.
This paper proposes a novel method based on first using a classifier which divides the perception-action space of the robot into subspaces and then generating a sperate model for each subspace. The viability of the proposed methodology has been demonstrated by teaching Scitos G5 mobile robot to achieve right wall following and complex route learning behaviours.