When a desired signal is applied to a servo system it responds in a characteristic fashion and follows the required trajectory with an error. The physical features of the actuators and the gain setting of the controller are the main parameters that determine the response of the system. Controllers with fixed gain values are effective for many conventional processes using slow-speed manipulators. However, there are several cases where the precise tracing of a fast trajectory under different payloads requires more advanced control techniques. When the motion is cyclical, learning control is one advanced technique which is appropriate to use. Depending solely on measurements of data from the preceding cycle, its implementation in real time is both fast and efficient. In practice, however, it has been observed that learning can induce high-frequency ripples on the tuned command curve which with increasing iterations result eventually in the saturation of the system's actuators. In this study, the use of on-line learning control techniques is discussed and a new approach using digital filters is implemented to prevent actuator saturation from occurring when learning is applied. A planar robotic manipulator has been designed and built to investigate the practical problems of learning control, particularly when the system runs at high speeds.
Memristive neural network for on-line learning and tracking with brain-inspired spike timing dependent plasticity