UCLA

Machine learning-based motion control of 3D characters is a classic computer graphicsproblem with applications spanning from the entertainment industry to robotics. In this thesis, we present novel approaches to achieving automatic and adaptive control of the simulated character motions that are anatomically consistent. More specifically, our main contribution is twofold. First, we develop a neuromuscular controller for a muscle-driven face-head-neck biomechanical model, which comprises 72 neck muscles and 52 facial muscles. Our pipeline transfers facial expressions and head poses from reference images and videos onto the biomechanical model in real-time. Second, we propose a deep reinforcement learning-driven controller for fish behavior in order to simulate ecologically valid underwater scenes. Our simulation framework, named DeepFoids, autonomously generates fish schooling patterns that accommodate environmental changes in factors such as neighbor interaction, light intensity, and underwater temperature. Our approaches are advantageous relative to previous methods as they (1) eliminate human labor in data collection, (2) require minimal manual calibration under different simulation settings, (3) retain a high level of biological accuracy, and (4) leverage the power of deep learning to capture the non-linear relations between the controlling variables and desired motions.