Evolving Balancing Controllers for Biped Characters in Games

Abstract

This paper compares two approaches to physics based, balancing systems, for 3D biped characters that can react to dynamic environments. The first approach, based on the concept of proprioception, use a neuro-controller to define the position and orientation of the joints involved in the motion. The second approach use a self-adaptive Proportional Derivative (PD) controller along with a neural network. Both neural networks were trained using a Genetic Algorithm (GA). The study showed that both approaches were capable of achieving balance and the GA proved to work well as a search strategy for both the neuro-controller and the PD-controller. The results also showed that the neuro-controller performed better but the PD-controller was more flexible and capable to recover under external disturbances such as wind drag and momentary collisions with objects.