Evolution of a Neural Network for Gait Animation

C. Hickey, C. Jacob, and B. Wyvill (Canada)


: gait control, genetic algorithm, neural networkcontroller, physically based animation


In this paper we describe efforts to create a physi cally based system that automatically produces realistic real-time animations of walking figures, controlled by a neural network, the weights and functionality of which is evolved by genetic algorithm techniques. In traditional computer graphics, the animator is forced to use intuition about the physical world in speci fying the motions of objects in a scene, but manual control techniques have generally proven to be unsatisfactory for realism. The use of dynamics greatly improves motion re alism, and shifts control of the animation from specifying absolute positions of objects to applying forces and torques to the objects in the scene. This is a much more difficult task for a human to control, whereas a neural network is well suited for this task. Manually programming a neural network is imprac tical, and training the network by back propagation is not straightforward, since there is no functional solution to be solved. Through the use of a genetic algorithm, one can determine the performance of a neural network as a whole, and select for whatever behavior is desired. Animator con trol is then directed through model design and behavior choices.

Important Links:

Go Back