ABSTRACT
Recent developments in neural network control at the Stanford Aerospace Robotics Laboratory are presented. A “Fully-Connected Architecture” (FCA) is developed for use with backpropagation (BP). This FCA has functionality beyond that of a layered network, and these capabilities are shown to be particularly beneficial for control tasks. A complexity control method is successfully used to manage the extra connections provided, and prevent over-fitting.
Second, a technique that extends BP learning to discontinuous functions is presented and applied to a difficult on-off thruster control problem. This method has many applications, namely any time a gradient-based optimization is used for systems with discontinuous functions. The modification to BP is very small simply requiring replacement of discontinuities with continuous approximations and injection of noise on the forward sweep.
The viability of both of these neural network developments is demonstrated by applying them to a thruster mapping problem characteristic of space robots. Real-world applicability is shown via an experimental demonstration on a 2-D laboratory model of a free-flying space robot.
