How To Simulate the World?
Speaker: Rishit Dagli
Abstract
I believe humans naturally perceive the world around them as rich, dynamic 3D scenes. To give learned models similar abilities, there are many ways to parameterize and several approaches that pursue this goal. I will share a quick overview of a small, admittedly biased subset of popular methods that: (1) use 3D Gaussian splats and dynamic 3D Gaussian splats for representation, (2) enable models to learn aspects of individual 3D objects and scenes, and (3) simulate them in a physically accurate way. I will also briefly discuss the types of priors commonly used, along with the explicit and implicit structures involved. I plan to leave plenty of time to discuss these aspects further.
Relevant Links
- Talk recording
- Slide deck
- Rishit Dagli's website
- Rishit Dagli's blog
- Squeeze3D: Your 3D Generation Model is Secretly an Extreme Neural Compressor
- VoMP: Predicting Volumetric Mechanical Property Fields
- Can Vision-Language Models Answer Face to Face Questions in the Real-World?
- SEE-2-SOUND: Zero-Shot Spatial Environment-to-Spatial Sound
- AirLetters: An Open Video Dataset of Characters Drawn in the Air
- torch-diffsim: A minimal differentiable physics simulator built entirely in PyTorch