Bumpy triangle meshes

[sparkprime]

I updated Bullet to svn revision 1785. I didn't notice any difference with or without setContactProcessingThreshold(0) when sliding a box over a btBVHTriangleMesh. It still jerks up into the air for most of the edges that it slides over. Sometimes the jerk is so violent the box does a flip.

[Erwin Coumans]

There are two separate problems:

1) A capsule shape or sphere shape should slide smoothly when using setContactProcessingThreshold(0). Please confirm or create a reproduction case in one of the Bullet demos, showing problems with capsule/sphere sliding over a triangle mesh.

2) A box, convex or gimpact shape sliding over a triangle mesh, hitting internal triangle edged. This requires fixing the triangle normal and penetration depth, using a contact callback, as mentioned in this issue. Have you applied this fix?
If it still doesn't work after applying this fix, please provide a reproduction case.

Thanks a lot,
Erwin

[sparkprime]

Ah setContactProcessingThreshold(0) is only for spheres and capsules... I never had any problem with spheres and have never used capsules. The fix you link to is actually written by me and I've been using it for over a year It does fix about 75% of the problem but there's still some catching going on. I can get it down to a reasonable level by just changing the step size though.

[gjaegy]

Hi,
sorry for resuscitating an old topic. I am having a similar issue when using the raycast vehicle. when the car is upside-down (after a crash), it slides on the ground but bumps when sliding on some edges.
I am going to apply the fix you mentioned Erwin. However I wonder why you don't fix this in the release directly ?

Cheers,
Gregory

[sparkprime]

The fix has a number of issues.

Firstly, objects can get 'jammed' in the ground, not necessarily between triangles. The bodies linear velocity slowly increases but their position never changes, not even by tiny amounts.

Secondly, if you have two triangles at a sharp angle to each other, e.g. if you have a wall (essentially a box, but using triangles) and a box sitting on top of the wall, it can end up toppling off and swinging inside the wall as it falls.

[sparkprime]

This is a video to demonstrate the second problem: http://www.youtube.com/watch?v=aQNiVzEWXD4

[gjaegy]

Understood Sparkprime.

So basically there is not perfect solution to this problem yet, is that right ?

[sparkprime]

If there is, I'd like to know about it

[Erwin Coumans]

There are some ways to improve the solution given in the issue tracker.

Sparkprime, could you help creating a reproduction case for each problem?
Thanks a lot,
Erwin

[vturchenko]

I just sent similar case to RP. Sphere falling perfectly down on perfectly horizontal mesh floor in VehicleDemo and bouncing sideways. Somebody explained to me before that it happens because Bullet treats triangles as convex hulls. I'm not sure how it could cause that but this behavior looks to me as a big flaw in Bullet system. Can that easily be fixed?

[Erwin Coumans]

I just sent similar case to RP. Sphere falling perfectly down on perfectly horizontal mesh floor in VehicleDemo and bouncing sideways. Somebody explained to me before that it happens because Bullet treats triangles as convex hulls. I'm not sure how it could cause that but this behavior looks to me as a big flaw in Bullet system. Can that easily be fixed?

If you want the sphere to collide using the triangle normals, just apply this fix.

Hope this helps.
Thanks,
Erwin

[kester]

I found when using the triangle normal that the contact point had to be reprojected. Otherwise the contact would break since the movement orthogonal to (new) normal was too large:

Code: Select all

/// Changes a btManifoldPoint collision normal to the normal from the mesh.
static void FixMeshNormal(btManifoldPoint& cp, const btCollisionObject* colObj0, int /*partId0*/, int /*index0*/)
{
	const btTriangleShape* tri_shape = static_cast<const btTriangleShape*>(colObj0->getCollisionShape());
	btVector3 tri_normal;
	tri_shape->calcNormal(cp.m_normalWorldOnB);
	cp.m_normalWorldOnB = colObj0->getWorldTransform().getBasis() * cp.m_normalWorldOnB;

	// Reproject collision point along normal.
	cp.m_positionWorldOnB = cp.m_positionWorldOnA - cp.m_normalWorldOnB * cp.m_distance1;
	cp.m_localPointB = colObj0->getWorldTransform().invXform(cp.m_positionWorldOnB);
}

[sparkprime]

kester that may fix my problem, i'll try it (hopefully soon)

[tp]

I have a situation where my game level is composed of tiles made of convex hulls that are placed right next to each other (most of my level is flat, i.e. the tops of the hulls are flat and they are at the same height). Both a capsule and a cylinder, when moved on top of the board, jump around when hitting tile edges, the cylinder doing considerably worse.

Does the issue described in this thread apply to adjacent convex hulls as well as adjacent triangles in a triangle mesh? Can I use the same callback mechanism to try and fix it?

It seems to me that convex hull faces have normals that are assumed to point directly perpendicular to the face, but edges have normals that are the average of the faces adjacent to that edge. Is this conclusion correct?

[sparkprime]

It wouldn't suprise me if you have a similar issue, but the contact_added hack won't work because it is designed for triangle meshes only. Hulls don't have triangles afaik so it would be harder to override the normal in these cases. You are also straddling two entire objects instead of just two subparts of a single object, but the way Bullet is designed I don't think this matters.