UCLA

This dissertation explores the frontier of biomimetic virtual human animation. Applying state-of-the-art machine learning techniques, we develop neuromuscular control frameworks that significantly enhance the naturalness and realism of simulated human motions. In particular, we address the challenges of achieving locomotion-based animation with biomechanical musculoskeletal fidelity and in the transfer of head pose and facial expressions from images and videos to a muscle-actuated model of the face-head-neck biomechanical complex. Key technical contributions of the thesis include (1) integration of Central Pattern Generator (CPG) controllers with a biomechanical body model, which enables the generation of adaptive and flexible locomotion patterns through reinforcement learning-based control parameter optimization, and (2) leveraging of the Facial Action Coding System in a computer-vision-based estimation of facial expressions and their transfer to a 3D face model via the coordinated activation of the muscles of facial expression. With regard to implementation, the migration of our simulation environments to Nvidia's GPU-accelerated Omniverse platform affords improved computational performance and advanced rendering techniques. By demonstrating the versatility of machine learning applied to muscle-driven face and body animation, this work advances the exploitation of biomimetic, physics-based human modeling and simulation in computer graphics and vision.