This research will develop a hybrid technique for the intelligent control of dexterous robotic manipulators. Tasks initially specified in a high-level task language will be encoded and refined in a neural-nework form. Phase I will determine the feasibility of applying this technique to the difficult problem of learning to position the loaded endpoint of a redundant manipulator subject to multiple goals and constraints. Rule- based control components will train, then shift control to, neural network components. Network performance will then be refined through reinforcement learning and on-line optimization. This technique, combining design convenience with implementation efficiency, will lay the foundation for robotic systems able to acquire skills for performing complex tasks in unstructured environment.
