Abstract by Andy Hall
Physically based learning for directable simulation in animation
Physical simulation is a boon for generating animations at scales and levels of detail that are prohibitive with manual techniques. However, an emphasis on physical accuracy in simulation methods can be at odds with the expressive and stylistic freedoms that art directors require, so methods that marry these generative benefits with intuitive means of direction and editing are highly desirable. Applying deep learning to this problem is an active and fruitful area of research, but normally, getting good results requires high volumes of training data. Hand-crafting example sets is a natural fit for artistic direction workflows, but not such a scale. Motivated by similar constraints on data availability, scientific researchers have been having success reducing the demand for data by augmenting deep learning models with prior physical laws as biases or constraints. This talk will thus explore possible directions for applying and extending physically based machine learning as a means of opening simulation methods to direction and control for producing detailed and visually plausible animations.