Robot motion planning and control using neural networks

Abstract

A method for motion path selection and control using a neural network that is especially applicable for manipulators with compliant limbs is proposed. The `trainability' of the neural network allows the design of path planners and controllers that have smooth force/torque-time profiles. For nonlinear control problems, performance is much better if the overall control function is decomposed into an acceleration controller neural network that outputs the required acceleration, and a linearizer neural network which compensates for nonlinearities. The feasibility of our method has been verified through simulations for both linear and nonlinear problems of one degree-of-freedom. Point-to-point control and smooth motion has been achieved with some overshoot.