Optimization and improvement of a robotics gaze control system using LSTM networks

Abstract

Gaze control represents an important issue in the interaction between a robot and humans. Specifically, deciding who to pay attention to in a multi-party conversation is one way to improve the naturalness of a robot in human-robot interaction. This control can be carried out by means of two different models that receive the stimuli produced by the participants in an interaction, either an on-center off-surround competitive network or a recurrent neural network. A system based on a competitive neural network is able to decide who to look at with a smooth transition in the focus of attention when significant changes in stimuli occur. An important aspect in this process is the configuration of the different parameters of such neural network. The weights of the different stimuli have to be computed to achieve human-like behavior. This article explains how these weights can be obtained by solving an optimization problem. In addition, a new model using a recurrent neural network with LSTM layers is presented. This model uses the same set of stimuli but does not require its weighting. This new model is easier to train, avoiding manual configurations, and offers promising results in robot gaze control. The experiments carried out and some results are also presented.