Avaliando simulações físicas de um personagem controlado por rede neural na unity

Abstract

Character animation is an area that is directly related to the movement expression of the characters. This area receives an incentive from the gaming and film industries to create tools that produce realistic animations. Using physics adds realism to animations and the use of physics in characters has been possible with the development of physical motors and with the need to study motion to build better controllers. This work proposes an infrastructure of components to aid the process of optimization of physically simulated characters controlled by neural network. The components are responsible for generating new neural networks, evaluating physical simulations controlled by the neural networks generated and selecting the best ones in order to evolve towards a desired behavior. In addition to the set of components, two control methods that explore the use of neural networks are presented. The addition of penalty functions to the fitness function is also shown as an option to improve search space behavior. Finally, the difficulties of obtaining suitable controllers are discussed.