Abstract
A fixed-final-state linear optimal tracking problem can be solved by applying a compensation input to a steady-state optimal regulator, and the compensation input can be computed by specifying a desired trajectory and terminal conditions. This paper proposes a learning method in which the compensation input is modified to reduce the trajectory error by using the error obtained by each trial. It is shown that even when the dynamics of the real plant is different from its mathematical model, the trajectory error by this method is also reduced monotonously with the iterative procedure under a certain condition. As the proposed method does not have any differential or phase lead operations on the error, it is free from noises of this kind. Results of simulation and experiments on the trajectory control of a two-link arm are shown, and the applicability of this method to robot arms with nonlinearities and differences of dynamics from their mathematical models is examined.
