Softbody explodes depending on node positions

[RauchMatthias]

Hello,
there seems to be a problem with btDeformableNeoHookeanForce and btDeformableLinearElasticityForce. Depending on the position of the nodes OR the order of the nodes in the tetra, the softbody grows exponentially in Volume. I modified the "VolumetricDeformable.cpp" sample file to reproduce the issue using a softbody consisting of 4 nodes and a single tetra.

Replace the "create volumetric softbody"-part of the "initPhysics()" function with the following:

`
// create volumetric soft body
{
// btSoftBody* psb = btSoftBodyHelpers::CreateFromTetGenData(getDeformableDynamicsWorld()->getWorldInfo(),
// TetraCube::getElements(),
// 0,
// TetraCube::getNodes(),
// false, true, true);
//
const int numNodes = 4;
btAlignedObjectArray x;
x.resize(numNodes);

    //explodes
	x[0] = btVector3(0, 0, 0);
	x[1] = btVector3(1, 0, 0);
	x[2] = btVector3(0, 0, 1);
	x[3] = btVector3(0, 1, 0);
	//works
	//x[0] = btVector3(0, 1, 0);
	//x[1] = btVector3(1, 0, 0);
	//x[2] = btVector3(0, 0, 1);
	//x[3] = btVector3(0, 0, 0);
	

	btSoftBody* psb = new btSoftBody(&getDeformableDynamicsWorld()->getWorldInfo(), numNodes, &x[0], 0);
	

	psb->appendTetra(0, 1, 2, 3); //explodes
	//psb->appendTetra(3, 1, 2, 0);  //works

	for (int i = 0; i < psb->m_tetras.size(); ++i)
	{
		psb->appendLink(psb->m_tetras[i].m_n[0], psb->m_tetras[i].m_n[1], 0, true);
		psb->appendLink(psb->m_tetras[i].m_n[1], psb->m_tetras[i].m_n[2], 0, true);
		psb->appendLink(psb->m_tetras[i].m_n[2], psb->m_tetras[i].m_n[0], 0, true);
		psb->appendLink(psb->m_tetras[i].m_n[0], psb->m_tetras[i].m_n[3], 0, true);
		psb->appendLink(psb->m_tetras[i].m_n[1], psb->m_tetras[i].m_n[3], 0, true);
		psb->appendLink(psb->m_tetras[i].m_n[2], psb->m_tetras[i].m_n[3], 0, true);
	}

	psb->initializeDmInverse();
	psb->m_tetraScratches.resize(psb->m_tetras.size());
	psb->m_tetraScratchesTn.resize(psb->m_tetras.size());



    getDeformableDynamicsWorld()->addSoftBody(psb);
    psb->scale(btVector3(2, 2, 2));
    psb->translate(btVector3(0, 15, 0));
    psb->getCollisionShape()->setMargin(0.1);
    psb->setTotalMass(0.5);
    psb->m_cfg.kKHR = 1; // collision hardness with kinematic objects
    psb->m_cfg.kCHR = 1; // collision hardness with rigid body
	psb->m_cfg.kDF = 2;
    psb->m_cfg.collisions = btSoftBody::fCollision::SDF_RD;
    psb->m_cfg.collisions |= btSoftBody::fCollision::SDF_RDN;
	psb->m_sleepingThreshold = 0;
    btSoftBodyHelpers::generateBoundaryFaces(psb);
    btDeformableGravityForce* gravity_force =  new btDeformableGravityForce(gravity);
    getDeformableDynamicsWorld()->addForce(psb, gravity_force);
    m_forces.push_back(gravity_force);
    

	//btDeformableNeoHookeanForce* neohookean = new btDeformableNeoHookeanForce(20, 80, .01);
    btDeformableLinearElasticityForce* linearElasticity = new btDeformableLinearElasticityForce(100,100,0.01);
	m_linearElasticity = linearElasticity;
	getDeformableDynamicsWorld()->addForce(psb, linearElasticity);
    m_forces.push_back(linearElasticity);
}`

Use the following lines to switch between working and exploding behavior:
psb->appendTetra(0, 1, 2, 3); //explodes
//psb->appendTetra(3, 1, 2, 0); //works
Alternatively switch the position of node 0 and 3 before the softbody creation.

With btDeformableNeoHookeanForce it explodes in the air, with btDeformableLinearElasticityForce it explodes when it hits the ground. Either way, I suspect forces being applied in the opposite direction.
btDeformableMassSpringForce does not have this problem.

[erwincoumans]

Unfortunately, the deformable code is an unsupported internship project. Not more we can do then just turn it into a discussion.
If you, or someone else, has time/resources to fix it, please let us know.

[RauchMatthias]

Ok, I found the reason, in btSoftbody::initializeDmInverse() the formula for calculating the tetra volume is missing the fabs(). The variable is called m_element_measure.
But I think that this is a design choice. Apparently there is a convention to order vertices following the right-hand rule, so that the determinant is always positive. The cube and bunny meshes follow this convention, I looked at it.
So, programmatically building a softbody works by calculating the determinant for each desired tetra, and if it is negative, switching any 2 of the nodes to make it positive. Then use btSoftbody::appendTetra.
For my application this is ok.
Therefore, if this is intended, there is no issue except maybe for the missing documentation. But at least it made me learn something.