The goal of this research is to establish a new method for nonlinear compliance control using neural nets to explore new possibilities in machine learning and control for robots and telemanipulators. Here, compliance is treated as nonlinear mapping from a measured force to a corrected motion and is represented by a multi-layer neural network, as well as by Gaussian networks. The objectives of the proposed research are three-fold. One is to develop a new method for representing "compliance" to deal with highly nonlinear, such as by stiffness and damping matrices. The second objective is to develop a learning methods for the generation and teaching of compliance, or force feedback strategies. The neural network approach allows us to teach a desired compliance from teaching data acquired from a human operator. It does not need explicit feedback laws and detailed task models such as those required for conventional analytic methods. It is hoped that the new approach will also allow us to transfer human skill in compliant motion control to robots and telemanipulators. The third objective of the proposed project is to develop a real-time, neural net controller that is involved directly in the feedback loop of robot control system. Efficient methods must be developed to analyze and design the nonlinear feedback system in order to accomplish smooth, stable responses. To achieve these goals, three subprojects will be conducted. Their objectives are to: 1. Establish a theoretical basis for the neural net representation of nonlinear compliance. 2. Develop techniques for acquiring teaching data, Processing the data, and training a neural network for learning a compliance, and 3. Develop a design method for real time, neural net feedback control in order to accomplish desirable dynamic response.
