Neural network sliding mode robot control

This paper develops a method for neural network control design with sliding modes in which robustness is inherent. Neural network control is formulated to become a class of variable structure (VSS) control. Sliding modes are used to determine best values for parameters in neural network learning rules, thereby robustness in learning control can be improved. A switching manifold is prescribed and the phase trajectory is demanded to satisfy both, the reaching condition and the sliding condition for sliding modes. A major objective of the work described has been to develop neural network architectures which will provide fast and robust on-line learning of the dynamic relations required by robot controller at the executive hierarchical level. The approach to proposed robot control involves using a neural network feedforward loop together with a discrete time 'chattering-free' feedback loop. Such a use of the neural network with a sliding mode learning algorithm is considered to be a new approach to adaptive control of a non-linear robot system. The advantage of the proposed control scheme prevails over those conventional model based control scheme since no precise knowledge of mathematical model is necessary. The algorithm was verified by experiments where inverted pendulum with the additional mass-spring-damper load was used.