Computer Graphics Laboratory ETH Zurich

ETH
CGL » Research » Simulation and Animation

Simulation and Animation

Introduction

Computer-animated objects are ubiquitous in entertainment and training applications of computer graphics (e.g. videogames, feature films, surgical simulators, etc.). As opposed to tedious and rather inflexible key-frame animation, physics-based simulation offers a concise, but rich and flexible way of defining the behavior of animated objects, by allowing the laws of physics to determine or guide their motion. In contrast to the mathematical modeling of physical objects in computational physics, our primary concerns are artistic controllability, robustness and efficiency. You can find more information on the simulation and animation group website.

Topics

Deep learning based physics simulations

We develop methods at the intersection of deep learning and physics simulations to solve some of the persistent problems in graphics. This includes our work on reduced order modeling and latent space integration, neural flow field recovery, and video-based reconstruction of flow fields.


Art-directable simulations with AI

In visual effects it is a key requirement to have art-directable numerical simulations and hence full control over the shape and motion of a body. We develop deep learning based methods to transfer styles from arbitrary input images to 3D smoke simulations, to compute smoke volumes and animations based on sketch inputs, and to control the motion of a fluid using keyframe inputs.


Physics-based digital humans

In the context of digital humans we develop techniques to model and animate faces using data and physics priors, applied to both visual effects and medical applications. This includes our work on reconstructing and simulating 3D faces for medical treatment planning, and on controlling physics-based facial animation.


Neural Augmentation of Simulation Representations of Robotic Systems

Digital twins that represent robotic systems with minimal sim-to-real gaps enable improved control and design. Within the scope of this effort, we augment simulation representations with neural networks to close sim-to-real gaps when modeling building blocks robotic systems are made of. Applications include surgical devices that require high-precision modeling, legged systems where accurate contact estimation is needed, and entertainment robots that can perform expressive animations.


Publications





© CGL, ETH Zurich