Bullet: Compound Shape Collision Fix

[Riordan-DC]

Background:
Kinematic body collisions resolved with the bullet physics space add together penetration recovery vectors from all shape collisions. The issue is that multiple shapes with the same penetration recovery vector push a kinematic body multiple times when only one push was required. The issue is outlined nicely in this post to the bullet forum in 2014: https://pybullet.org/Bullet/phpBB3/viewtopic.php?t=10843

(taken from the bullet forum post from user: Enhex)

FIX:
This fix is to offset the collision shape by the recovery vector before detecting penetration so that repeat pushes are not detected. The change reverses the lines in this commit which added compound collision support:
6dd65c0#diff-28f7bca21ebbe699cd16802c05d43f4e085331c7d0427c38da0ab3e6d14dc696L1186
6dd65c0#diff-28f7bca21ebbe699cd16802c05d43f4e085331c7d0427c38da0ab3e6d14dc696L1217

I admit i have no idea the consequences for such a change but i assume they should be confined to KinematicBody.

This hopefully fixes a lot of bullet physics jitter issues seen here (from @pouleyKetchoupp bullet issue tracker):
#29392
#34596

Test without fix:

without_fix-converted.mp4

Test with fix:

with_fix-converted.mp4

Bugsquad edit:

• Fixes [Bullet] RigidBody2D bounces when passing on top of two perfectly aligned StaticBody2Ds #28438
• Fixes [Bullet] move_and_slide_with_snap causes erroneous movement near edges #29392
• Fixes [Bullet] KinematicBody: move_and_slide_with_snap intersection edges problem #34436
• Fixes [Bullet] Jittering against two discrete, connected walls #34596
• Fixes [Bullet] GridMap collisions broke from 3.1.2 to 3.2 #35713
• Fixes [Bullet] KinematicBody jitters up and down when walking along a flat gridmap. #35780

[akien-mga]

Could you provide a test project that demonstrates the bug and can be used to confirm the fix?
It would then also be interesting to see how it behaves with GodotPhysics.

It would also be good to test this change against https://github.com/godotengine/godot-demo-projects/tree/master/3d/physics_tests (well whichever tests actually use the modified code) to ensure that it performs as expected. I'll try to test it myself but if you want to do some extra testing, that's very welcome.

For the record for potential testers: this is best tested against the 3.x branch (the patch should be easy to apply) since the Bullet module is currently disabled in master.

[Riordan-DC]

@akien-mga
This test demonstrates the bug. It walks two kinematic bodies towards two static body walls. One wall has 1 collision shape and the other has 4 shapes. On 3.4.2 with Bullet physics the kinematic body that collides with the 4 shape wall will jitter while the other wont.

This demonstration shows this issue isnt present in GodotPhysics.

Running this demonstration with this patch the bullet behavior should match the correct GodotPhysics behavior.

test.zip

EDIT: I don't have much experience with the 3D Physics Tests but at a glance the Bullet physics behavior for a kinematic body on a moving platform seems identical between 3.4.2 and this patch.

[akien-mga]

Sorry for the delay. We've removed the Bullet backend in master since it was disabled for more than a year, and our aim for 4.0 is to have it provided as a GDExtension plugin instead of built-in.

It's still used in 3.x and this PR seems relevant to merge there, could you redo it against the 3.x branch?

[Riordan-DC]

Sorry for the delay. We've removed the Bullet backend in master since it was disabled for more than a year, and our aim for 4.0 is to have it provided as a GDExtension plugin instead of built-in.

It's still used in 3.x and this PR seems relevant to merge there, could you redo it against the 3.x branch?

Sure thing.
Does that require making a new PR against 3.x or can I change this one?
Also I'm very glad we are still supporting Bullet for now :)

[akien-mga]

It might be easiest to make a new PR targeting the 3.x branch. Otherwise you can indeed change the base branch for this PR as you had attempted (which then tries to merge master + your commit on 3.x, hence a ton of commits), and then you need to force push an update with the commit on top of your local 3.x branch.

Otherwise I can do it myself too so we keep the useful discussion above in this PR.

[akien-mga]

Now it should be good to merge, sorry for the delay :)

[Riordan-DC]

Thanks so much for your help getting this merged. :))

[akien-mga]

The change reverses the lines in this commit which added compound collision support:
6dd65c0#diff-28f7bca21ebbe699cd16802c05d43f4e085331c7d0427c38da0ab3e6d14dc696L1186
6dd65c0#diff-28f7bca21ebbe699cd16802c05d43f4e085331c7d0427c38da0ab3e6d14dc696L1217

This is worth highlight for review. This change was done in #27415 together with a lot of other fixes, but I don't see it discussed specifically. @madmiraal did identify this PR as the cause for some of the regressions linked below (#34596 (comment)).

This hopefully fixes a lot of bullet physics jitter issues seen here (from @pouleyKetchoupp bullet issue tracker):
#29392
#34596

I tested the MRPs from those two issues respectively in 3.5 RC 1 and this PR and confirmed that this PR fixes them 👍

• [Bullet] Jittering against two discrete, connected walls #34596: Does seem to be a regression from 27415 which removed those two lines.
• [Bullet] move_and_slide_with_snap causes erroneous movement near edges #29392: Actually predates 27415 (merged in 3.2 and this is reported against 3.1), but there seems to have been a bug introduced in 3.1 that affected both GodotPhysics and Bullet. This was fixed in Properly pass safe margin on initialization. #40377 but it was found to only solve the bug for GodotPhysics, despite being implemented in KinematicBody itself - this was likely due to the 27415 change that is being reverted here.

This also seems to fix the following issues which are linked in the above issues or reference #27415:

• [Bullet] RigidBody2D bounces when passing on top of two perfectly aligned StaticBody2Ds #28438.
• [Bullet] KinematicBody jitters up and down when walking along a flat gridmap. #35780.
• [Bullet] GridMap collisions broke from 3.1.2 to 3.2 #35713.

There might be more, still testing. I'll update when I've reviewed everything I can find.

Some other related issues which do not seem fixed as of this PR (no change):

• [Bullet] move_and_slide() causes jitter when colliding with corners between walls #32182: Tested all 3 MRPs linked in that issue.

These might warrant further debugging and might not be related to the 3.1 and 3.2 regressions discussed above.

[akien-mga]

Thanks a lot, this was a long awaited fix!

[akien-mga]

I found one issue where behavior seems to have regressed with this fix: #36334.

With the MRP in that issue, enabling "Sync To Physics" on the moving KinematicBody platform as suggested, there's now back and forth on is_on_floor when the platform is moving upwards by being animated.

Here's the updated MRP from that issue which prints is_on_floor(): moving_platform.zip

---

I also noticed a behavior change with the MRP in #45058. It's buggy before, and buggy after but in a weird different way where the player cube disappears.

[elvisish]

#35780 is still not fixed if the KinemeticBody is moving slow enough, it'll get stuck on gaps in gridmaps (both ground and wall colliders). It does not do this with Godot Physics. It only seems to happen if gravity is being applied to ground, or if the KB is walking into a wall while sliding along it.