Stochastic Methods for Neural Network Robot Control-Silicon Valley Campus - Carnegie Mellon University

Stochastic Methods for Neural Network Robot Control

robot movement
Figure 1: Robot learning to move in noisy environment.

Over the last two decades significant advances have been made in the neural network control of robot manipulators for numerous applications. However, to date, only limited progress has been made when considering a stochastic element in the closed loop system. These types of stochastic models are becoming increasingly important as the size of the physical robots continues to shrink. We are currently pursuing an extension of existing adaptive and neural network control algorithms to account for various noise sources. In particular, we are analyzing the system in a stochastic framework in which the Ito lemma of stochastic calculus is frequently used to determine the stochastic stability of the closed loop system. Additionally, we have been pursuing the use of simulated annealing for neural network control in which the stochastic disturbance can be shown to enhance learning in an unknown environment.

Contact: Abe Ishihara