Deep Neuro-Evolution: Evolving Neural Network for Character Locomotion Controller

Designing the controller to actuate the virtual character is challenging given the high dimensions of both character states and controls. Flexible user control is normally required in interactive applications (virtual and augmented reality), which further adds to constructing a motion controller for the virtual characters. Although deep neural networks showed their potential in tackling this problem, existing works based on neural networks normally require the manual design of the network architecture, which requires professional experience and a trial-and-error process. This paper proposes an automatic solution to optimize both the topology and connection weights of the deep neural network, used as a character locomotion controller. This method mimics the evolution process in nature, evaluates individual fitness, and generates the individuals for the next generation while maintaining the current generation's fitness level. We show that this method is convenient to construct the locomotion controller for virtual characters on a general basis.