← ArchivePaper2019
Subspace Neural Physics: Fast Data-Driven Interactive Simulation
Daniel Holden, Bang Chi Duong, Sayantan Datta, Derek Nowrouzezahrai
Neural network trained in a learned subspace to replace expensive physics simulation with fast interactive deformation for characters.
Abstract
Data-driven methods for physical simulation are an attractive option for interactive applications due to their ability to trade precomputation and memory footprint in exchange for improved runtime performance. Yet, existing data-driven methods fall short of the extreme memory and performance constraints imposed by modern interactive applications like AAA games and virtual reality. Here, performance budgets for physics simulation range from tens to hundreds of micro-seconds per frame, per object. We present a data-driven physical simulation method that meets these constraints. Our method combines subspace simulation techniques with machine learning which, when coupled, enables a very efficient subspace-only physics simulation that supports interactions with external objects - a longstanding challenge for existing subspace techniques. We also present an interpretation of our method as a special case of subspace Verlet integration, where we apply machine learning to efficiently approximate the physical forces of the system directly in the subspace. We propose several practical solutions required to make effective use of such a model, including a novel training methodology required for prediction stability, and a GPU-friendly subspace decompression algorithm to accelerate rendering.
How to read this
- Category
- Method: data-driven neural physics for interactive deformation
- Contributions
- A neural network operating in a learned subspace to approximate physical forces and replace costly full simulation
- Support for interaction with external objects, a long-standing limitation of subspace techniques
- An interpretation as subspace Verlet integration, plus a training methodology to make the model usable under tight memory and time budgets
- Context
- Relates to data-driven deformation approximation (referenced Fast and Deep Deformation Approximations, Bailey 2018) and to classical subspace simulation, coupling the two with machine learning.Builds on: Fast and Deep Deformation Approximations
- Correctness
- Targets the tens-to-hundreds-of-microseconds-per-object budgets of games and VR and is presented as an approximation that trades precomputation and memory for runtime; being subspace and data-driven, accuracy is bounded by the learned basis and training distribution, so out-of-distribution motions and interactions are a natural risk.
- Clarity
- Reasonably accessible with a helpful Verlet-integration framing; a first pass conveys the subspace-plus-learning idea, a second pass clarifies the force approximation and training methodology.
- How to read it
- Read for the subspace-Verlet interpretation and the external-interaction handling first; second pass on the training methodology and the practical solutions section if you intend to hit interactive budgets yourself.
Builds on
Related work
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →