Neural Learning of Flexible Full Body Motion
Neural learning of motor behavior has been a long standing topic at the border of neuroscience and robotics and there are many efforts to find the neural correlates underlying motor control to inspire formal neural network models. The project will explore the power of a recent approach to recurrent networks -- reservoir computing -- in motor learning. Motivated by preliminary studies which demonstrate that inverse models of the highly redundant robot ASIMO can be learned in this neurally inspired fashion, the project aims at learning task prediction and motor models in a single framework and completely data driven. In a developmental and imitation learning perspective most conventional schemes are neither flexible nor fast enough to cope with varying parameters like tool use (which changes the end-effector position) or configuration changes of the robot. One way to overcome that problem is to let the robot perform exploratory movements, from which a fully self-learning model can be established. Which movements to explore in the infinite dimensional space of body trajectories and the determination of efficient strategies for acquisition of a corresponding basic motion repertoire is a further major goal of the project.