Neuroevolution for vision-based seeking behavior in 3D soft voxel robots

Abstract

This paper details a simple experiment that tests two genetic encodings, NEAT and HyperNEAT, for the evolution of vision-based food-seeking behavior in neural-controlled 3D soft voxel robots. The evolution of food-seeking behavior is a preliminary step towards ecosystems of advanced artificial animals, in which the animals seek both food and mates. Two environments were tested: with and without deadly obstacles. Traditional evolutionary search was used, with an objective-based fitness function. Both NEAT and HyperNEAT encodings were tested for the evolution of robot neural controllers. The results of the experiment showed the NEAT encoding resulted in increasingly effective food-seeking behavior over time, whereas experiments with the HyperNEAT encoding did not achieve the desired behavior. This suggests that NEAT at least is a viable algorithm to evolve neural networks for the task of vision-based object-seeking in complex robots, and warrants further experimentation. On the other hand, HyperNEAT struggled with this task. This could be due to a number of reasons, including a common issue like EA being stuck in local optima, or because the encoding might struggle to evolve and represent irregular structures required for the task.