Robot Control in Dynamic Environments Using Memory-Based Learning

Abstract

Learning systems used on robots typically require a-priori knowledge in the form of environmental models or trial-and-error approaches requiring a robot to physically execute multitudes of trial solutions. Neither approach is very suitable for dynamic unstructured environments in which a robot is sent to explore and gather information prior to human entry. This chapter presents a new approach, ‘memory-based learning’, in which a robot is provided with an initial baseline behavior whose performance is linked with a metric explicitly defined by a function whose arguments are sensory inputs and resulting robot actions. A neural network, using sensor inputs and action outputs, has been chosen as the basic controller building block for behaviors as it is very amendable to rapid in-situ generation and adaptation. The use of a neural network controller for a maneuvering task using electronic ‘ears’ for sensory input is demonstrated in this work, and genetic algorithms are shown to be an effective method for rapidly developing the network weights and transfer functions. The concept of memory based learning is introduced and a construct for representing coupled sense/action information is presented and demonstrated using both simple reactive and memory based controllers, whose performance is demonstrated through simulation studies and on land robots.