Adaptive neural network control and learning for robot manipulator

A new adaptive neural control approach is proposed by using Radial basis function (RBF) network for the robot manipulator with completely unknown parameters. In previous adaptive neural control, the problem of whether adaptive neural controllers indeed learn the unknown system dynamics has less been investigated. For dissatisfying the persistent excitation (PE) condition, the convergence of neural weights to their optimal values can not be guaranteed, as a consequence, the adaptive neural controller has to be retrained redundantly even for repeating the same control task, which may waste time and energy. The designed adaptive neural controller not only achieves uniformly ultimately boundness of all signals in the closed-loop system, but also achieves the convergence of partial neural weights and locally-accurate approximation of unknown closed-loop system dynamics along periodic or recurrent tracking orbit, i.e., deterministic learning. The learned knowledge represented in a time-invariant and spatially distributed manner and stored as constant neural weights can be used to improve control performance, and can also be recalled and reused in the same or similar control task, so that the robot can be easily controlled with little effort. Simulation studies are included to demonstrate the effectiveness of the approach. ©2013 Journal of Mechanical Engineering.