Abstract
This paper studies the performance of both P-type and PI-type learning schemes for robot motion control. P-type learning is one of the simplest scheme that does not use differentiation of velocity signals. Recent theoretical results show that a forgetting factor in P-type learning guarantees the robustness against initialization errors, fluctuation of dynamics, and measurement noises. In this paper, experiments are carried out to confirm the theory, and design guidelines of learning control parameters are clarified on the basis of the experimental results. Moreover, a new PI-type learning algorithm that uses positional signals is proposed, and the uniform boundedness of output signal is shown theoretically. The proposed PI-type learning scheme improves the learning speed and reduces trajectory errors considerably in comparison with the P-type learning scheme.
