WIP: Geometry Nodes: playing with PBD #129718

Jacques Lucke · 2024-11-02T14:29:26+01:00

Jacques Lucke commented

2024-11-02 14:29:26 +01:00

This is an experiment to get to know position based dynamics (PBD) better.

This adds a PBD Solver node which implements some basic constraints for softbody, hair and grain simulation.

Information about PBD:

This is an experiment to get to know position based dynamics (PBD) better. This adds a PBD Solver node which implements some basic constraints for softbody, hair and grain simulation. Information about PBD: * https://animation.rwth-aachen.de/media/papers/2015-EG-Tutorial.pdf * https://www.youtube.com/watch?v=qISgdDhdCro&list=PLfd9K5jHLvZC1dE_O68lc3Jf0kUvidYFr <video src="https://projects.blender.org/attachments/2ef22c22-1938-4b53-91e6-f2ac53b13bbc" controls />

sim_test.blend

1.1 MiB

2024-11-02 14-38-40.mov

1.0 MiB

👍 2 👀 10 ❤️ 8 🎉 3 🚀 5

Jacques Lucke added 5 commits 2024-11-02 14:29:36 +01:00

add initial solver node a2499b2d25

initial softbody and hair solver a18aaacf37

Merge branch 'main' into pdb-solver-experiment 6732aa413b

initial point point collisions 77609cdaf6

progress bae1d09122

Jacques Lucke added 1 commit 2024-11-02 14:38:15 +01:00

improve floor friction a02c1fbe4e

Hans Goudey changed title from ~~WIP: Geometry Nodes: playing with PDB~~ to WIP: Geometry Nodes: playing with PBD

2024-11-02 14:50:36 +01:00

Iliya Katushenock added this to the Nodes & Physics project 2024-11-02 16:52:22 +01:00

CGmatter commented

2024-11-02 19:06:50 +01:00

First-time contributor

highly recommend throwing the 'x' in there lol (for xpbd)
all you need is to add a tiny term in the constraint that is dependent on dt^2

3 main insights ive learned when playing with this:

surprisingly you get better/faster convergence spending all your resources on substeps (while maintaining a single constraint solve): https://matthias-research.github.io/pages/publications/smallsteps.pdf
using xpbd kills any stiffness-dependence based on substeps/constraint steps - this is good because you want your simulation to look very similar regardless of the number of steps: https://matthias-research.github.io/pages/publications/XPBD.pdf
verlet integration is super cheap and stable

literally read everything by matthias - he's the G

highly recommend throwing the 'x' in there lol (for xpbd) all you need is to add a tiny term in the constraint that is dependent on dt^2 3 main insights ive learned when playing with this: --- 1) surprisingly you get better/faster convergence spending all your resources on substeps (while maintaining a single constraint solve): https://matthias-research.github.io/pages/publications/smallsteps.pdf 2) using xpbd kills any stiffness-dependence based on substeps/constraint steps - this is good because you want your simulation to look very similar regardless of the number of steps: https://matthias-research.github.io/pages/publications/XPBD.pdf 3) verlet integration is super cheap and stable literally read everything by matthias - he's the G

👍 5

Jacques Lucke commented

2024-11-02 20:57:07 +01:00

Thanks. I'm aware of the "x" part (I'm kind of using the term PBD for everything related to that currently). Still have to figure out how everything fits together though :)

🚀 3

seanterelle commented

2024-11-16 00:00:02 +01:00

First-time contributor

Love to see XPBD getting a tighter integration, thanks for working on this!

One thing I appreciate about implementing xpbd within node groups rather than having a dedicated solver node is the control one maintains over the simulation.

In my experiments with xpbd in Blender, the main missing piece is an efficient way of obtaining contact points between bodies.

In my opinion a contact points node would close the gap for more efficient collision response for simple bodies such as spheres, and could implement GJK and EPA to get contact points between convex bodies.

It's already possible to create xpbd rigid body simulations in geo nodes, but it requires using a repeat zone to iterate through possible colliders and detect contact points.

A collision points node that outputs a new "collision points" geometry could be used with the attribute statistic node to make that process efficient, and would be a great tool to have in addition to dedicated solvers!

edit: I may have spoken too soon, because you would need a group_id input to the attribute statistic node for this to work. It should be possible currently with the accumulate field node though.

Love to see XPBD getting a tighter integration, thanks for working on this! One thing I appreciate about implementing xpbd within node groups rather than having a dedicated solver node is the control one maintains over the simulation. In my experiments with xpbd in Blender, the main missing piece is an efficient way of obtaining contact points between bodies. In my opinion a contact points node would close the gap for more efficient collision response for simple bodies such as spheres, and could implement GJK and EPA to get contact points between convex bodies. It's already possible to create xpbd rigid body simulations in geo nodes, but it requires using a repeat zone to iterate through possible colliders and detect contact points. A collision points node that outputs a new "collision points" geometry could be used with the attribute statistic node to make that process efficient, and would be a great tool to have in addition to dedicated solvers! edit: I may have spoken too soon, because you would need a group_id input to the attribute statistic node for this to work. It should be possible currently with the accumulate field node though.

👍 3

Damian commented

2024-11-21 01:37:27 +01:00

First-time contributor

Sorry if this is not the place for that type of question:
Is this an experimental branch of Blender?
Because this is what I see in 4.3 and 4.4 when opening the .blend file from the header post of this thread:

Sorry if this is not the place for that type of question: Is this an experimental branch of Blender? Because this is what I see in 4.3 and 4.4 when opening the .blend file from the header post of this thread: <img width="450" alt="image" src="attachments/5c16e49f-b1b1-4214-82bd-ebc5ecdbe0c0">

image.png

52 KiB

Jacques Lucke commented

2024-11-21 12:21:08 +01:00

This is a personal experiment of mine, it's not really something that can be tested well by others unfortunately, because I hard-coded all kinds of assumptions in the code.

Will close this to avoid confusion.

This is a personal experiment of mine, it's not really something that can be tested well by others unfortunately, because I hard-coded all kinds of assumptions in the code. Will close this to avoid confusion.

👍 1

Jacques Lucke closed this pull request

2024-11-21 12:21:10 +01:00

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

WIP: Geometry Nodes: playing with PBD #129718

Pull request closed