Goal-Directed Movement Generation with a Transient-Based Recurrent Neural Network Controller

Abstract

We introduce a control framework based on a recurrent neural network for goal-directed movement generation. We exploit the network dynamics to implement a nonlinear task space controller. Efficient online learning and execution of the network makes the proposed approach adaptive and real-time capable. We achieve reliable and excellent generalization for the 7-DOF redundant PA-10 robot arm in simulation.