Teaching and learning of compliance using neural nets: representation and generation of nonlinear compliance

A new approach to the representation and learning of compliance control laws using neural networks is presented. Compliance is treated as a nonlinear mapping from a measured force to a corrected motion. The nonlinear mapping is represented by a multilayered neural network, which makes it possible to deal with complex strategies of force feedback. This network representation provides not only linear compliance, as represented by stiffness and damping matrices, but also nonlinear compliance, which has never been explored extensively. This network approach also makes it possible to teach a desired compliance from teaching data by using an iterative learning algorithm. Conventional methods of compliant motion control are reviewed, linear compliance and its limits are discussed, and nonlinear compliance based on the network representation is introduced. An analysis is made of the nonlinearity of compliance that arises in performing assembly tasks, and the network structure that meets the requirements for representing a group of nonlinear compliances necessary for performing the assembly tasks is obtained