z80
commented
1 year ago Hello NVidia,
After debugging the phenomena for a moment it feels that the problem lays in multiplying a quaternion by a very close to it inverted quaternion in two places in /physx/source/lowleveldynamics/src/DyFeatherstoneForwardDynamic.cpp. It feels that due to finite float32 data type precision sometimes the result instead of being close to (x=0, y=0, z=0, w=1) happens to become close to (x=0, y=0, z=0, w=-1).
The suggestion is instead of
if (newParentToChild.w < 0.f)
do
if ( (newParentToChild.w * relativeQuat.w) < 0.f )
It looks like it fixes the problem. Please find the patch for "105.1-physx-5.3.0" tag pasted below.
diff --git a/physx/source/lowleveldynamics/src/DyFeatherstoneForwardDynamic.cpp b/physx/source/lowleveldynamics/src/DyFeatherstoneForwardDynamic.cpp
index 7e479af8..dd20926a 100644
--- a/physx/source/lowleveldynamics/src/DyFeatherstoneForwardDynamic.cpp
+++ b/physx/source/lowleveldynamics/src/DyFeatherstoneForwardDynamic.cpp
@@ -1946,7 +1946,8 @@ namespace Dy
const PxU32 dofs)
{
PxQuat newParentToChild = (newRot.getConjugate() * pBody2WorldRot).getNormalized();
- if(newParentToChild.w < 0.f)
+ //if(newParentToChild.w < 0.f)
+ if ( (newParentToChild.w * relativeQuat.w) < 0.f )
newParentToChild = -newParentToChild;
//PxQuat newParentToChild = (newRot * pBody2WorldRot.getConjugate()).getNormalized();
@@ -1966,11 +1967,13 @@ namespace Dy
return newParentToChild;
}
- PxQuat computeSphericalJointPositions(const PxQuat& /*relativeQuat*/,
- const PxQuat& newRot, const PxQuat& pBody2WorldRot)
+ //PxQuat computeSphericalJointPositions(const PxQuat& /*relativeQuat*/,
+ PxQuat computeSphericalJointPositions(const PxQuat& relativeQuat,
+ const PxQuat& newRot, const PxQuat& pBody2WorldRot)
{
PxQuat newParentToChild = (newRot.getConjugate() * pBody2WorldRot).getNormalized();
- if (newParentToChild.w < 0.f)
+ //if (newParentToChild.w < 0.f)
+ if ( (newParentToChild.w * relativeQuat.w) < 0.f)
newParentToChild = -newParentToChild;
/*PxQuat jointRotation = newParentToChild * relativeQuat.getConjugate();
jcarius-nv
preist-nvidia
Hello!
I'm working with PxArticulationGeneralizedCoordinate. No forces are applied, gravity is zero, no collisions are created. Some of the joint angles instantly change after a single scene->simulate() call. I've tested several PhysX versions: PhysX 4, PhysX 5.1.3 and PhysX 5.3.0 on Windows 10 and Ubuntu Linux 18.04. The behavior is identical in all of them.
The issue happens only with a specific armature topology provided in the snippet code below and works as expected with different topologies.
All links have mass of 1kg and identity inertia matrix. All joints are either eSPHERICAL or eREVOLUTE. No control is applied. I expect all angles to remain zero. But angles 17, 18, 19 change instantly. Please see the example application attached (for PhysX 5.1.3) and the result output "image.png".
Thank you!
Library and Version
PhysX v5.1.3
Operating System
Windows 10, Ubuntu Linux 18.04.
Steps to Trigger Behavior
Code Snippet to Reproduce Behavior
Expected Behavior
When created, all joint angles are zero. Environment takes a single time step of dt=1.0/240.0 seconds. Due to no forces are applied, gravity is zero, no collision volumes are created, and due to all mass and inertia are assigned to be 1, it is expected that all joint angles remain exactly zero or be negligibly small.
Actual Behavior
After a single call to scene->simulate( 1.0/240.0 ) and reading articulation->copyInternalStateToCache( *cache, physx::PxArticulationCacheFlag::ePOSITION | physx::PxArticulationCacheFlag::eVELOCITY ); cache->jontPosition[17] is -5.43, cache->jontPosition[18] is 3.11, cache->jontPosition[19] is -0.6.