Bitte wählen Sie ihr Lieferland und ihre Kundengruppe
The work characterizes the Cerebellar Model Articulation Controller (CMAC) neural network and develops a hybrid control system merging the trainable CMAC neural network with the well-structured knowledge-based control architecture Real-Time Control System (RCS) creating an easily programmed sensory interactive controller. The work provides techniques for tuning the CMAC parameters to store functions of interest. Previous RCS control systems were analyzed to develop a standard technique for merging CMAC processing with the knowledge-based state table processing of the control modules. This was tested in a simple RCS structure using a CMAC module to carry out an object avoidance maneuver. The CMAC was trained to develop the correct output responses using real data from ten pairs of infrared proximity emitter/detector pairs dealing with significant sensor characteristic mismatch, considerable signal noise, and large variations in reflected signal due to surface characteristics and angle. The work led to the design and fabrication of a 3-axis experimental robot arm to be used with the CMAC processing of sensor data in more complex activities. The arm was constructed with small gear ratios to allow other factors such as inertial and Coriolis forces in addition to friction to have a measurable effect on performance.