Deterministic learning from robust adaptive NN control of robot manipulators

摘要

We investigate the learning issue in the robust adaptive neural-network (NN) control process of manipulator with unknown system dynamics and disturbance. Based on recently developed deterministic theory, the regression vector of an appropriately designed robust adaptive NN controller satisfies the partial persistent exciting (PE) condition when tracking a periodic or periodic-like reference orbit, and partial convergence of NN weights can be achieved when the disturbance term is small in the training phase, the dynamics of manipulators can be learned by NN in the form of constant neural weights. When repeating the same or similar control tasks, even if the disturbance is large, the learned knowledge can still be recalled and reused to achieve guaranteed stability and better control performance with little effort, tremendous repeated training phase of NN can be avoided. Both of the true learning and the reusing of learned knowledge are realized.