sweepTest uses incorrect AABB

[Dragonlord]

Code in question:

Code: Select all

void	btCollisionWorld::convexSweepTest(const btConvexShape* castShape, const btTransform& convexFromWorld, const btTransform& convexToWorld, ConvexResultCallback& resultCallback, btScalar allowedCcdPenetration) const
{

	BT_PROFILE("convexSweepTest");
	/// use the broadphase to accelerate the search for objects, based on their aabb
	/// and for each object with ray-aabb overlap, perform an exact ray test
	/// unfortunately the implementation for rayTest and convexSweepTest duplicated, albeit practically identical



	btTransform	convexFromTrans,convexToTrans;
	convexFromTrans = convexFromWorld;
	convexToTrans = convexToWorld;
	btVector3 castShapeAabbMin, castShapeAabbMax;
	/* Compute AABB that encompasses angular movement */
	{
		btVector3 linVel, angVel;
		btTransformUtil::calculateVelocity (convexFromTrans, convexToTrans, 1.0f, linVel, angVel);
		btVector3 zeroLinVel;
		zeroLinVel.setValue(0,0,0);
		btTransform R;
		R.setIdentity ();
		R.setRotation (convexFromTrans.getRotation());
		castShape->calculateTemporalAabb (R, zeroLinVel, angVel, 1.0f, castShapeAabbMin, castShapeAabbMax);
	}

#ifndef USE_BRUTEFORCE_RAYBROADPHASE

	btSingleSweepCallback	convexCB(castShape,convexFromWorld,convexToWorld,this,resultCallback,allowedCcdPenetration);

	m_broadphasePairCache->rayTest(convexFromTrans.getOrigin(),convexToTrans.getOrigin(),convexCB,castShapeAabbMin,castShapeAabbMax);
(...)

The castShapeAabbMin and castShapeAabbMax is incorrect. I've slamed my head against the walls for days trying to figure out why a small sphere (used to replace the non-working rayHits test) moving along a long ray (say 5m or so) does not hit a set of objects. Then I stumbled over that piece of code. As you can see the AABB is calculated from the sphere AABB (which is small) using angular rotation... but without linear velocity. The AABB used to test against the broadphase only composes of the shape itself. The distance travelled is not added to the AABB causing every sweepTest going farther than the basic shape AABB is going to miss geometry.

Now to add insult to injury look at the next piece of code:

Code: Select all

void	btCollisionWorld::rayTest(const btVector3& rayFromWorld, const btVector3& rayToWorld, RayResultCallback& resultCallback) const
{
	//BT_PROFILE("rayTest");
	/// use the broadphase to accelerate the search for objects, based on their aabb
	/// and for each object with ray-aabb overlap, perform an exact ray test
	btSingleRayCallback rayCB(rayFromWorld,rayToWorld,this,resultCallback);

#ifndef USE_BRUTEFORCE_RAYBROADPHASE
	m_broadphasePairCache->rayTest(rayFromWorld,rayToWorld,rayCB);
#else
	for (int i=0;i<this->getNumCollisionObjects();i++)
	{
		rayCB.process(m_collisionObjects[i]->getBroadphaseHandle());
	}	
#endif //USE_BRUTEFORCE_RAYBROADPHASE

}

Looks not too suspicious unless you look at rayTest. It has 3 parameters but the function definition has 5 parameters:

Code: Select all

virtual void	rayTest(const btVector3& rayFrom,const btVector3& rayTo, btBroadphaseRayCallback& rayCallback, const btVector3& aabbMin=btVector3(0,0,0), const btVector3& aabbMax = btVector3(0,0,0)) = 0;

It has two default parameters for the AABB which default to zero vectors. Hence if you test a ray of lets say 5m length the broadphase is only tested at the orign of the ray ignoring the entire length to the end point. This explains also why ray tests most of the time fail. The AABB simply fails to detect anything not including the ray origin.

These findings are all on a 2.82 code base. I've seen no bug tracker so I'll drop it in here.