Evolving motion control for a modular robot

This paper documents our ongoing efforts in devising efficient strategies in motion control of the brittle star-typed robot. As part of the control framework, each robotic leg consisting of series of homogenous modules is modeled as a neural network. The modules representative of neurons are interconnected via synaptic weights. The principle operation of the module involves summing the weighted input stimulus and using a sinusoidal activation function to determine the next phase angle. Motion is achieved by propagating phase information from the modules closest to the main body to the remainder of the modules in the leg via the synaptic weights. Genetic algorithm was used to evolve near optimal control parameters. Simulations results indicate that the current neural network inspired control model produces better motion characteristics than the previous cellular automata-based control model as well as addresses other issues such as fault tolerance.