Generation and Organization of Behaviors for Autonomous Robots

Doctoral thesis, 2006

In this thesis, the generation and organization of behaviors for autonomous robots is studied within the framework of behavior-based robotics (BBR). Several different behavioral architectures have been considered in applications involving both bipedal and wheeled robots. In the case of bipedal robots, generalized finite-state machines (GFSMs) were used for generating a smooth gait for a (simulated) five-link bipedal model, constrained to move in the sagittal plane. In addition, robust balancing was achieved, even in the presence of perturbations. Furthermore, in simulations of a three-dimensional bipedal robot, gaits were generated using clusters of central pattern generators (CPGs) connected via a feedback network. A third architecture, namely a recurrent neural network (RNN), was used for generating several behaviors in a simulated, one-legged hopping robot. In all cases, evolutionary algorithms (EAs) were used for optimizing the behaviors. The important problem of behavioral organization has been studied using the utility function (UF) method, in which behavior selection is obtained through evolutionary optimization of utility functions that provide a common currency for the comparison of behaviors. In general, the UF method requires the use of simulations. Thus, an important part of this thesis has been the development of a general-purpose software library (UFLib) implementing the UF method. In order to study the properties of the UF method, several behavioral organization problems, mostly involving wheeled robots, have been considered. Most importantly, it was found that the UF method greatly simplifies the search for solutions to a wide variety of behavioral organization problems and requires a minimum of hand-coding. Furthermore, the results show that the use of multiple simulations (for the evaluation of a robot) significantly improves the ability of the robot to select appropriate behaviors. For the EA, it was found that the standard crossover procedure, which swaps entire utility functions between individuals, performed at least as well as several modified operators, and that the mutation rate should be set so as to generate around three parameter modifications per individual. Finally, some early results are presented concerning the use of the UF method in connection with a robot intended for transportation and delivery.