Neural-Network Estimation of the Variable Plant for Adaptive Sliding-Mode Controller

The Lyapunov based theoretical development of a neural-network sliding-mode based estimation of highly non-linear and variable robot plant for a direct-drive robot controller is shown in the paper. Derived adaptive control law was tested for four types of robot neural-network sliding-mode controllers: centralized, simplified centralized, decentralized and simplified decentralized, which were verified on a real laboratory direct-drive 3 D.O.F. PUMA like mechanism. Centralized and decentralized control approaches estimate only a part of the variable robot dynamic model (torque model due to friction, Coriolis, centripetal and centrifugal forces) and use only the part of a dynamic plant model (the so called estimated inertia matrix M). Both simplified methods do not need any plant model parameter for an accurate estimation of the direct-drive robot plant, but need some more time to learn dynamic model parameters. All four types of the neural network continuous sliding-mode controllers were successfully tested for algorithm's adaptation capability for sudden changes in the manipulator dynamics (load).