rapier3d convex hull regression from 0.15 to 0.16

[LPGhatguy]

Hello! We recently upgraded from rapier3d 0.15 to 0.16. We use rapier's built-in convex hull generation on some of our meshes to get a decent default collision mesh.

We have a few test scenes with dynamic bodies. One of them is a series of barrels sitting on tables that the player can shoot at. The barrel colliders are using convex hull colliders, generated by feeding all of the model's vertices into ColliderBuilder::convex_hull.

In rapier3d 0.15.0, those barrels sit on the table and are stable:

rapier3d-0.15.0.mp4

When upgrading to rapier3d 0.16.0, the barrels roll off of the table when the simulation starts:

rapier3d-0.16.0.mp4

The model is one of Kenney's free assets. Here is a zip file containing the glTF 2.0 file we load: barrel.zip.

The collisions generated from this model weren't perfect in 0.15.0, but they have regressed in 0.16.0. I understand that the data that I'm giving rapier may not be great. Using convex decomposition instead results in much better collisions but takes a long time. Is there anything else I can submit to help diagnose this issue?

[sebcrozet]

Hi! Could you by any chance integrate Rapier’s debug-renderer, so I can understand exactly the geometry of the scene?

[LPGhatguy]

Hey! Yeah, I think I could probably do that. How do I do that? I've been putting off building our own physics visualization, but if there's something built-in, that sounds great!

[sebcrozet]

First, you need to implement the DebugRenderBackend trait. All that is required is to be able to render a colored line. Then you can create a DebugRenderPipeline and call its render method at each frame.

[LPGhatguy]

Wow, that debug renderer interface is amazing, I wish I had integrated it sooner!

Here is rapier3d 0.15.0: https://i.eryn.io/2246/rapier3d-0.15.0-debug.mp4

Here is rapier3d 0.16.1: https://i.eryn.io/2246/rapier3d-0.16.1-debug.mp4

[rfurlan]

I believe I have another example of this issue. A mesh associated with a convex hull collision proxy is dropped a few units above a plane and comes to rest atop of it on 0.15.0:

However on 0.16 and up, it behaves differently after initial contact with the plane.

1 frame after contact:

5 frames after contact:

10 frames after contact:

15 frames after contact:

On frame 16, the remaining visible portion of mesh associated with the rigid body instantly tunnels through the plane:

[rfurlan]

The convex hull itself was generated correctly:

[sebcrozet]

@LPGhatguy, @rfurlan Thank you both for the additional details. Could you please add the following to your Cargo.toml and see if that fixes the regression you are observing ?

[patch.crates-io]
parry3d = { git = "https://github.com/dimforge/parry", branch = "odd-convex" }

[rfurlan]

@sebcrozet yes everything behaves as expected with the patch, thank you :)