Neural networks for gesture-based remote control of a mobile robot

We present a neural network architecture for gesture-based interaction between a mobile robot and its user, thereby spanning a bridge from the localisation of the user over the recognition of its gestural instruction to the generation of the appropriate robot behavior. Since this system is applied under real-world conditions, especially concerning the localisation of a human user, some proper techniques are needed which have an adequate robustness. Hence, the combination of several components of saliency towards a multi-cue approach, integrating structure- and color-based features, is proposed. At the moment, the gestures themselves are very simple and can be described by the spatial relation between face and hands of the person. The organisation of the appropriate robot behavior is realised by means of a mixture of neural agents, responsible for certain aspects of the navigation task. Due to the complexity of the whole system, above all we use "standard neural network models", which are modified or extended according to the task at hand. Preliminary results show the reliability of the overall approach as well as the sufficient functionality of the already realised sub-modules.