UCLA

Physical simulation has long been a cornerstone of computer graphics, driving the generation of vivid dynamics across a variety of forms, from movies to video games. In recent years, the development of machine learning algorithms, along with the rise of massively parallel hardware such as GPUs, has created new opportunities as well as new challenges for traditional physical simulation. For instance, mobile devices such as smartphones and AR/VR glasses demand high-speed simulations on resource-constrained hardware. Similarly, the fields of embodied AI and robotics require fast and robust physical solvers capable of handling complex interactions. Moreover, advances in modern vision techniques and generative models have opened new avenues for using physical simulation to create novel digital content.

In this dissertation, we first present novel reduced-order modeling techniques to accelerate existing physical simulation methods. Next, we introduce a robust rigid-deformable simulation method tailored for robotic applications. Finally, we explore how bridging physical simulation with state-of-the-art vision techniques enables dynamic novel view synthesis, and how combining physical simulation with 3D generative models facilitates the creation of 3D assets that stably interact with gravity, contact, and friction.