Deep learning for motion synthesis

In the attempt of making cheap, high quality motion synthesis systems, the research community is developing algorithms for automatically controlling agents to perform a given desired high level task like walking or jumping. These algorithms are referred to as motion control systems. The goal of this master's thesis was to apply supervised deep learning techniques to solve this motion control problem for two particular tasks: making a biped to walk and keeping balance of a 3D human character under random external forces. Agent control data was collected using the C-PBP algorithm and the goal was to create deep learning networks that could replicate C-PBP performance in the former two tasks. It is shown that this approach fulfills its goal for simple tasks like the biped problem. However, curse of dimensionality limits its capacity and prevents it to solve more challenging problems like the 3D character problem. In addition, some experiments with LBN stochastic neural network architecture were conducted in order to get more information on how they work. These experiments showed some of LBNs problems and limitations during training