Robust force/motion control of flexible-joint robots using neural networks

A general framework for the force/motion control of flexible-joint robot is proposed. Many conventional robust adaptive schemes can be utilized to control the system. In this paper, we present one novel neural network (NN) method which does not require the robot dynamics to be exactly known. Compared with adaptive control, no linearity in the unknown parameters is needed and no persistency of excitation condition is required. Compared with other NN approaches, our method does not require an off-line "training phase". All errors including force, position and weight are guaranteed to be bounded. The main advantage of such a framework is that robustness to parametric uncertainties is achieved without the availability of acceleration and jerk measurement. The complexity of the controller is at the same level as the rigid joint case.