Comparison and combination of learning controllers - Computational enhancement and experiments

Five discrete-frequency linear learning-control laws are compared and experimentally tested. These include simple integral-control-based learning using a single learning gain, phase-cancellation learning control, a contractionmapping learning-control law with monotonic decay of the error norm, and learning controllers that invert the system model and the observer model. The inversion designs converge the fastest initially, but phase cancellation with identification updates and the contraction-mapping method with model updates have better stability robustness properties. The learning control approaches are combined to obtain the advantages of each, by using inversion methods for the first few repetitions, followed by a more robust method. It is demonstrated that the computation of the learning action can be made in the frequency domain using fast Fourier transform methods, with as much as 94% reduction in computarion time. In experiments on a Robotics Research Corporation robot, the learningcontrol laws result in a reduced rms tracking errors for all joints, by a factor of close to 3 orders of magnitude.