Dirtying Jacobian update flags whenever the parent transform is changed -- Release 5.0 patch

[mxgrey]

This is a bug fix and regression test for release-5.0 which fixes issue #499.

Thanks to Nathan Ratliff for bringing this issue to our attention!

[jslee02]

I agree with that the transform of a parent BodyNode can be changed without its child BodyNodes being informed that their Jacobians are now dirty. Suppose we have a serial chain and the position of i-th joint changed then only the child BodyNodes Jacobian will be dirty when all the Jacobians of descendant BodyNodes should be dirty as well. In that sense, it seems good solution to set the flags at here instead of Joint::notifyPositionUpdate().

One concern is that we might want to update this comment a bit. If I'm correct, the Jacobian of a BodyNode is actually dependent on the parent's Jacobian and the relative transform (or the joint positions of the parent Joint) of the BodyNode (to transform the coordinates of the parent's Jacobian from the parent's to the BodyNode's).

Also, it would be good to clarify more why and when these flags must be dirtied regardless of whether the relative transform of this BodyNode is already dirty. Basically, we need this for the case, when BodyNode::notifyTransformUpdate() is called, Jacobian is updated while the relative transform of this BodyNode is not updated yet. However, this case wouldn't happen if the BodyNode has a parent BodyNode. The relative transform will be consequently updated when the Jacobian is updated to be used in transforming the coordinates of the parent's Jacobian. You can check this in BodyNode::updateBodyJacobian(). That being said, we place blew code here (before the relative transform's dirtiness check) for the case this BodyNode doesn't have parent BodyNode. Maybe this might be too much detailed note but it would prevent us from spending time to interpret again the intention of this code in the future.

[mxgrey]

Suppose we have a serial chain and the position of i-th joint changed then only the child BodyNodes Jacobian will be dirty when all the Jacobians of descendant BodyNodes should be dirty as well. In that sense, it seems good solution to set the flags at here instead of Joint::notifyPositionUpdate().

This is a good point. It may be the case that the regression test is only passing because of the fact that we request the transform of the final BodyNode before dirtying the Jacobian. I'm going to adjust the regression test to account for this, and I bet it's going to start failing.

[mxgrey]

I'm now thinking we might need two update tracks:

notifyJacobianUpdate() would only deal with updating the Jacobians, and it would cut off at whichever BodyNode already knows its Jacobian is dirty (if a parent BodyNode's Jacobian is dirty, then all its children's Jacobians must be dirty as well). notifyTransformUpdate() will remain the same, but the very first thing it will do is call its own notifyJacobianUpdate().

However, notifyJacobianUpdate() will need to be a member of the JacobianNode class, not the BodyNode class. That way all JacobianNode types (such as EndEffector) will make use of it.

[jslee02]

This sounds good to me! These updates might be able to be applied to osgDart branch or master once the branch is merged since JacobianNode is introduced by osgDart branch.

In addition, we might need notifyJacobianDerivUpdate(), and it should be called by BodyNode::notifyVelocityUpdate()

[mxgrey]

I've made the changes that have been discussed, and I've reworked the unittest so that it fails without these latest changes.

[jslee02]

I created PR #503 based on the above discussion.