Motivation and Local Image Entropy Based Measures in Evolutionary Mobile Robot Navigation

Abstract

Robotic navigation in unknown and unstructured environments is a complex and difficult task fraught with potential problems and peril to the navigator. Extensive research has been made and is ongoing into this problem with the aid of various robotic architectures, sensors and processors. Behavior based robotics is an approach that in general does not promote the use of complex world models or symbolic knowledge. This design philosophy promotes the idea that robots should be low cost, built incrementally, capable of withstanding sensor and other noise, and without complex computers and communication systems. Behavior based learning systems typically include reinforcement learning, neural networks, genetic algorithms, fuzzy systems, case and memory based learning (1). These biologically based mechanisms are capable of novel complex behaviors which avoid local minima and have the ability to extrapolate from training information. One area of research in behavior based robotics has focused on providing more natural and intuitive interfaces between robots and people (2; 3). One recent investigation (4), decouples specific robot behavior using an intuitive interface based on biological motivations (e.g. curiosity, hunger, etc) (5). Training the robot to behave according to said motivations requires optimization of the robot inference system (e.g. neural network) which in our approach is implemented using a genetic algorithm (GA). It is a well known fact in machine learning (6), that having diversity during training can provide for the emergence of more robust systems which are capable of coping with a variety of environmental challenges (7; 8). Early studies have shown that information theory can be used as an aid in analyzing robotic learning performance in terms of the diversity of information received during training (4; 9). Performing a more extensive analysis of environment training diversity using such information theoretic based measures was something of interest to us. Toward this goal we investigate the capability of the entropy based environmental and motivation measures toward analyzing the outcome of several robotic navigation experiments. In our work, robot navigation is performed in a simulator (10) by providing sensor values directly into a neural network inference engine that drives left and right motors. The robot uses infrared sensors which give limited information about the surroundings in which the robot is located. In order to reduce the complexity of the action space, action-based environmental modeling (AEM) (11) is implemented with a small action set of four basic actions (e.g. go straight, turn left, turn right, turn around) in order to encode a sequence of actions based 22