It has been difficult to develop robust, adaptive, goal- directed robotic behaviors based on sensory inputs, and to reprogramm robotic systems to carry out new tasks. This study will develop a hybrid control system that merges a knowledge- based control architecture with trainable neural-network modules. The control structure will handle task control and data management while the neural-net modules process sensor data and control actuator servos. Research will characterize relationships between network parameters and storable function types, speed of training, accuracy, and time delay effects; define boundaries between neural processing and more conventional task representation; and define methods of extracting model descriptions from trained neural nets. Demonstration tasks employing hardware implementations will be used to validate results. A prototype of a testbed hybrid control system is planned for Phase II, with Phase III development of the testbed into a low-cost, trainable, PC-based tool set to allow fast programming specification of complex sensorimotor tasks.
