Robot Behavior Learning with a Dynamically Adaptive RBF Network: Experiments in Offline and Online Learning

In this paper a dynamically adaptive neural network architecture is investigated for robot behavior learning. Specifically, a so-called “Grow When Required” network (GWR) is used to dynamically cluster the sensor-motor training data for determining the centers of a radial basis function network (RBF), and then the RBF network is trained for acquiring and performing the required behaviors. We illustrated our learning system by making experiments from simple behaviors, such as wall following and obstacle avoiding, to some complicated behaviors, such as moving object following and path learning, which were conducted on a real robot. We also tested the learning system in online training mode. Experimental results showed that our learning system is able to learn a wide range of robot behaviors due to its dynamically adaptive learning structure.