Genetic Algorithms for Gait Synthesis in a Hexapod Robot M. Anthony Lewis, Andrew H. Fagg and George A. Bekey Institute for Robotics and Intelligent Systems and Center for Neural Engineering University of Southern California Los Angeles, CA 90089-0781, USA Abstract This paper describes the staged evolution of a complex motor pattern generator (CPG) for the control of the leg movements of a six-legged walking robot. The CPG is composed of a network of neurons. In contrast to the main stream work in neural networks, the interconnection weights are altered by a Genetic Algo- rithm (GA), rather than a learning algorithm. Staged evolution is used to improve the convergence rate of the algorithm, thus obtaining rapid evolution of behavior toward a goal set. First, an oscillator for the individual leg movements is evolved. Then, a network of these oscillators is evolved to coordinate the movements of the different legs. In this way, the designer specifies "islands of fitness" on the way to the final goal, rather than using a single fitness function or determining the ex- plicit solution to the control problem. By introducing a staged set of manageable challenges, the algorithm's performance is improved. These techniques may be applicable to other complex or ill-posed control prob- lems in robot control. The system itself determined how to evolve from one island to the next through the GA. 1. Introduction A unique feature in the design and synthesis of robotic walking machines is the need to develop a method for generation and control of the sequences of leg movements representing specific gait patterns. In many animals such sequences are produced by special neural structures known as central pattern generators (CPGs). For engineers designing walking machines, the gait sequence can be produced by an algorithm, a control law or a pre-programmed gait sequence. Differences between these approaches include the issues of representation of knowledge, robustness with respect to changes in the environment and the facility of the selected language to represent important aspects of the control problem. For example, continuous control laws are well suited for the representation of the dynamics of fast moving walking machines, but they may require knowledge of accurate dynamic models of their behavior. Finite state machines can more easily rep- resent the periodic sequences seen in all forms of animal and machine walking. Once a representation is selected, the engineer must translate the observed walking behavior into the ap- propriate code. Such a coded representation is a difficult task and, in general, not a solved problem. In this chapter we demonstrate an alternative approach to gait synthesis, using neural networks as the 'lan- Published in: Zheng, ed. Recent Trends in Mobile Robots, pp 317-331, World Scientific, New Jersey, 1994.