I've been getting back into sim/pbd stuff after nearly a decade, and there have been a couple really interesting Position Based Dynamics papers released recently that I wanted to put on everyone's radar:
This one is really mind-blowing -- tiny time steps with a single-solver-iteration per step behaves a lot better than bigger steps with multiple solver iterations: https://matthias-research.github.io/pag ... lsteps.pdf
And more recently, a fully PBD-based rigid body simulation: https://matthias-research.github.io/pag ... Bodies.pdf
it looks like they still haven't solved the "collision detection needs to be moved into the solver, which is a big bottleneck" problem -- they're still doing "speculative constraints" (generating collision constraints for all *nearby* objects and then during solving skipping any that don't penetrate), and at a lower rate than the tiny steps. I wonder if an incremental approach might help, since when you're ticking at 1-2.4khz things are barely moving each tick.
recent PBD papers
-
- Posts: 1
- Joined: Thu Feb 18, 2021 4:58 pm
Re: recent PBD papers
I'm a big fan of the Detailed Rigid Bodies paper for a few reasons.
For one, it explains the guts of rigid body simulation better than other papers in the field that typically reference a dense thesis or abstract mathematical paper. There's a lot of distilled wisdom in that paper which is great for engineering applications.
Secondly, the paper eliminates a lot of the concerns with PBD around using real-world units.
I don't know of any engines that are using the "many substeps, single iteration" method yet. Though they demonstrate very promising results with it.
For one, it explains the guts of rigid body simulation better than other papers in the field that typically reference a dense thesis or abstract mathematical paper. There's a lot of distilled wisdom in that paper which is great for engineering applications.
Secondly, the paper eliminates a lot of the concerns with PBD around using real-world units.
I don't know of any engines that are using the "many substeps, single iteration" method yet. Though they demonstrate very promising results with it.