The evolution of modular artificial neural networks for legged robot control

This paper outlines a system that allows a neural network, which is used to control a robot, to evolve in a structured but open-ended way. The final intention of the research is that, as the network develops, intelligence will eventually emerge. This is accomplished by placing the robot in a developing environment and allowing both this environment and the robot's body form, sensors and actuators to become more complex and sophisticated as time passes. As this development takes place, neural network modules are added to the control system. The result is that the robot's complexity and that of the neural network grows with its environment. Results are presented showing the system in operation on a simulated legged robot.