| This paper presents a neural network which learns to control a robotic manipulator with minimal a-priori knowledge of the system kinematics and dynamics. The controller is a back-propagation network with a novel inhibitory structure. Positional, velocity, acceleration, and goal vectors are normalized and fed forward to successive elements of the hierarchy. Simulations conducted on various network topologies and arm configurations suggest that state space representation can be satisfactorily learned from a set of different repeated trajectories by a neural network. |
