Sliding-mode approach for on-line neural identification of robotic manipulators

Feedforward neural networks are commonly used for online modeling, identification and adaptive control purposes in case variations in process dynamics or in disturbance characteristics are present. In this paper a novel variable-structure-systems-based learning algorithm is applied to on-line neural identification of robotic manipulators. The zero level set of the learning error variable is considered as a sliding surface in the neural identifier learning parameters space. The proposed learning approach represents a simple, yet robust mechanism for guaranteeing finite time reachability of zero learning error condition. Off-line optimization of the learning scheme configuration by a genetic algorithm is implemented in advance to achieve complexity reduction and performance improvement. The proposed neural identification scheme is experimentally tested on a CRS 255 industrial manipulator. The results show that the neural model inherits some of the advantages of the sliding mode control approach, such as high speed of learning and robustness, and is able to follow the actual robot joint trajectories with a high accuracy.