Acquiring Efficient Locomotion in a Simulated Quadruped through Evolving Random and Predefined Neural Networks

The acquisition and optimization of dynamically stable locomotion is important to engender fast and energy efficient locomotion in animals. Conventional optimization strategies tend to have difficulties in acquiring dynamically stable gaits in legged robots. In this paper, an evolving neural network (ENN) was implemented with the aim to optimize the locomotive behavior of a four-legged simulated robot. In the initial generation, individuals had neural networks (NNs) that were either predefined or randomly initialized. Additional investigations show that the efficiency of applying additional sensors to the simulated quadruped improved the performance of the ENN slightly. Promising results were seen in the evolutionary runs where the initial predefined NNs of the population contributed to slight movements of the limbs. This paper shows how a predefined ENNs linked to bio-inspired sensors can optimize a locomotive strategy for a simulated quadruped.