12/01/1989 09:00 AM
Computer Science
In the past, large amounts of user effort have been required to create acceptable animations in commonly used motion control systems for animation. In order to reduce user effort, research in computer animation has been directed towards automatic motion control systems. A new method of automatic motion control using neural networks (also known as connectionist models) is examined as one possible means of reducing user effort. The main objective of such a method is to train a neural network to control the inbetween motion given only the start and goal positions. If using neural networks is an appropriate method for automatic motion control, not only would it have applications in computer animation but robotics as well. In order to analyze the feasibility of using neural networks for automatic motion control, the simple motion of an arm reaching for an object in a two dimensional world was used. Initially, a multi-layered feed-forward network was trained to predict the position of an arm at the following time frame, given only the position of the arm at the current time frame and its relative position from the object. After the network was trained using the back-propagation learning algorithm, the performance of the network was tested, producing an animated motion that was far from acceptable, even though the network\'s training performance leveled off to a low mean error. It was also discovered that a highly acceptable animated motion could be produced when the network continued to be trained during the testing phase. Other network configurations and data representations were also explored to lower the mean error during training and to produce more promising arm motions. The results of this research indicate that many of the simpler techniques that have been suggested in the neural net literature simply don\'t work for automatic motion control. However, more sophisticated techniques may show some promise, even though the results of this research are inconclusive. Notes: M.S. Thesis
