A Study on Falling Cat Landing Problem Using Composite Neuroevolution

Abstract

Cat has ability to safely land in the free fall. In this research, we first construct a virtual cat robot by Physics Modeling. Then we use Composite Neuroevolution to optimize its controller. Composite Neuroevolution optimizes multiple Artificial Neural Networks which are integrated into a complete controller in a pre-defined way. On the contrast, common Neuroevolution algorithms, for example CMA-ES, NEAT, CNE, ESP, optimize one Artificial Neural Network. In our experiment, only Composite Neuroevolution can converge. The result shows the landing behavior is achieved by our method.